on Mathem 241
Surveys in Combinatorics, 1997 Edited by R. A. Bailey
CAMBRIDGE UNIVERSITY PRESS
LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES Managing Editor: Professor J.W.S. Cassels, Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, 16 Mill Lane, Cambridge CB2 ISB, England The titles below are available from booksellers, or, in case of difficulty, from Cambridge University Press. 46 50 59 66 69 86 87 88 89
90 92 93 96 97 98 99 100 104 105 107 109 113 114 116 118 119 121 128
129 130 131 132 133 134 135 137 138 139 140 141
144 146 148 149 150 151
152 153 155 156 158 159 160 161
162 163 164 166 168 169
170 171
172 173
p-adic analysis: a short course on recent work, N. KOBLITZ Commutator calculus and groups of homotopy classes, H.J. BAUES Applicable differential geometry, M. CRAMPIN & F.A.E. PIRANI Several complex variables and complex manifolds II, M.J. FIELD Representation theory, I.M. GELFAND et al Topological topics, I.M. JAMES (ed) Surveys in set theory, A.R.D. MATHIAS (ed) FPF ring theory, C. FAITH & S. PAGE An F-space sampler, N.J. KALTON, N.T. PECK & J.W. ROBERTS Polytopes and symmetry, S.A. ROBERTSON Representation of rings over skew fields, A.H. SCHOFIELD Aspects of topology, I.M. JAMES & E.H. KRONHEIMER (eds) Diophantine equations over function fields, R.C. MASON Varieties of constructive mathematics, D.S. BRIDGES & F. RICHMAN Localization in Noetherian rings, A.V. JATEGAONKAR Methods of differential geometry in algebraic topology, M. KAROUBI & C. LERUSTE Stopping time techniques for analysts and probabilists, L. EGGHE Elliptic structures on 3-manifolds, C.B. THOMAS A local spectral theory for closed operators, 1. ERDELYI & WANG SHENGWANG Compactification of Siegel moduli schemes, C: L. CHAI Diophantine analysis, J. LOXTON & A. VAN DER POORTEN (eds) Lectures on the asymptotic theory of ideals, D. REES Lectures on Bochner-Riesz means, K.M. DAVIS & Y: C. CHANG Representations of algebras, P.J. WEBB (ed) Skew linear groups, M. SHIRVANI & B. WEHRFRITZ Triangulated categories in the representation theory of finite-dimensional algebras, D. HAPPEL Proceedings of Groups - St Andrews 1985, E. ROBERTSON & C. CAMPBELL (eds) Descriptive set theory and the structure of sets of uniqueness, A.S. KECHRIS & A. LOUVEAU The subgroup structure of the finite classical groups, P.B. KLEIDMAN & M.W. LIEBECK Model theory and modules, M. PREST Algebraic, extremal & metric combinatorics, M: M. DEZA, P. FRANKL & I.G. ROSENBERG (eds) Whitehead groups of finite groups, ROBERT OLIVER Linear algebraic monoids, MOHAN S. PUTCHA Number theory and dynamical systems, M. DODSON & J. VICKERS (eds) Operator algebras and applications, 1, D. EVANS & M. TAKESAKI (eds) Analysis at Urbana, 1, E. BERKSON, T. PECK, & J. UHL (eds) Analysis at Urbana, 11, E. BERKSON, T. PECK, & J. UHL (eds) Advances in homotopy theory, S. SALAMON, B. STEER & W. SUTHERLAND (eds) Geometric aspects of Banach spaces, E.M. PEINADOR & A. RODES (eds) Surveys in combinatorics 1989, J. SIEMONS (ed) Introduction to uniform spaces, I.M. JAMES Cohen-Macaulay modules over Cohen-Macaulay rings, Y. YOSHINO Helices and vector bundles, A.N. RUDAKOV et a! Solitons, nonlinear evolution equations and inverse scattering, M. ABLOWITZ & P. CLARKSON Geometry of low-dimensional manifolds 1, S. DONALDSON & C.B. THOMAS (eds) Geometry of low-dimensional manifolds 2, S. DONALDSON & C.B. THOMAS (eds) Oligomorphic permutation groups, P. CAMERON L-functions and arithmetic, J. COATES & M.J. TAYLOR (eds) Classification theories of polarized varieties, TAKAO FUJITA Twistors in mathematics and physics, T.N. BAILEY & R.J. BASTON (eds) Geometry of Banach spaces, P.F.X. MULLER & W. SCHACHERMAYER (eds) Groups St Andrews 1989 volume 1, C.M. CAMPBELL & E.F. ROBERTSON (eds) Groups St Andrews 1989 volume 2, C.M. CAMPBELL & E.F. ROBERTSON (eds) Lectures on block theory, BURKHARD KULSHAMMER Harmonic analysis and representation theory, A. FIGA-TALAMANCA & C. NEBBIA Topics in varieties of group representations, S.M. VOVSI Quasi-symmetric designs, M.S. SHRIKANDE & S.S. SANE Surveys in combinatorics, 1991, A.D. KEEDWELL (ed) Representations of algebras, H. TACHIKAWA & S. BRENNER (eds) Boolean function complexity, M.S. PATERSON (ed) Manifolds with singularities and the Adams-Novikov spectral sequence, B. BOTVINNIK Squares, A.R. RAJWADE Algebraic varieties, GEORGE R. KEMPF Discrete groups and geometry, W.J. HARVEY & C. MACLACHLAN (eds)
174 175 176 177 178 179 180 181
182 183 184 185 186 187 188 189 190 191
192 193 194 195 196 197 198
199
200 201
202 203 204 205 206 207 208 209 210 211
212 214 215
216 217 218 220 221
222 223 224 225 226 227 228 229 230 231
232 233 234 235 236 237 238 239 240 241
242 243 244 245
Lectures on mechanics, J.E. MARSDEN Adams memorial symposium on algebraic topology 1, N. RAY & G. WALKER (eds) Adams memorial symposium on algebraic topology 2, N. RAY & G. WALKER (eds) Applications of categories in computer science, M. FOURMAN, P. JOHNSTONE & A. PITTS (eds) Lower K- and L-theory, A. RANICKI Complex projective geometry, G. ELLINGSRUD et al Lectures on ergodic theory and Pesin theory on compact manifolds, M. POLLICOTT Geometric group theory I, G.A. NIBLO & M.A. ROLLER (eds) Geometric group theory II, G.A. NIBLO & M.A. ROLLER (eds) Shintani zeta functions, A. YUKIE Arithmetical functions, W. SCHWARZ & J. SPILKER Representations of solvable groups, O. MANZ & T.R. WOLF Complexity: knots, colourings and counting, D.J.A. WELSH Surveys in combinatorics, 1993, K. WALKER (ed) Local analysis for the odd order theorem, H. BENDER & G. GLAUBERMAN Locally presentable and accessible categories, J. ADAMEK & J. ROSICKY Polynomial invariants of finite groups, D.J. BENSON Finite geometry and combinatorics, F. DE CLERCK et a! Symplectic geometry, D. SALAMON (ed) Computer algebra and differential equations, E. TOURNIER (ed) Independent random variables and rearrangement invariant spaces, M. BRAVERMAN Arithmetic of blowup algebras, WOLMER VASCONCELOS Microlocal analysis for differential operators, A. GRIGIS & J. SJOSTRAND Two-dimensional homotopy and combinatorial group theory, C. HOG-ANGELONI, W. METZLER & A.J. SIERADSKI (eds) The algebraic characterization of geometric 4-manifolds, J.A. HILLMAN Invariant potential theory in the unit ball of Cn, MANFRED STOLL The Grothendieck theory of dessins d'enfant, L. SCHNEPS (ed) Singularities, JEAN-PAUL BRASSELET (ed) The technique of pseudodifferential operators, H.O. CORDES Hochschild cohomology of von Neumann algebras, A. SINCLAIR & R. SMITH Combinatorial and geometric group theory, A.J. DUNCAN, N.D. GILBERT & J. HOWIE (eds) Ergodic theory and its connections with harmonic analysis, K. PETERSEN & I. SALAMA (eds) An introduction to noncommutative differential geometry and its physical applications, J. MADORE Groups of Lie type and their geometries, W.M. KANTOR & L. DI MARTINO (eds) Vector bundles in algebraic geometry, N.J. HITCHIN, P. NEWSTEAD & W.M. OXBURY (eds) Arithmetic of diagonal hypersurfaces over finite fields, F.Q. GOUVEA & N. YUI Hilbert C*-modules, E.C. LANCE Groups 93 Galway / St Andrews 1, C.M. CAMPBELL et a! Groups 93 Galway / St Andrews II, C.M. CAMPBELL et al Generalised Euler-Jacobi inversion formula and asymptotics beyond all orders, V. KOWALENKO, N.E. FRANKEL, M.L. GLASSER & T. TAUCHER S. DAVID (ed) Stochastic partial differential equations, A. ETHERIDGE (ed) Quadratic forms with applications to algebraic geometry and topology, A. PFISTER Surveys in combinatorics, 1995, PETER ROWLINSON (ed) Algebraic set theory, A. JOYAL & I. MOERDIJK Harmonic approximation, S.J. GARDINER Advances in linear logic, J.-Y. GIRARD, Y. LAFONT & L. REGNIER (eds) Analytic semigroups and semilinear initial boundary value problems, KAZUAKI TAIRA Computability, enumerability, unsolvability, S.B. COOPER, T.A. SLAMAN & S.S. WAINER (eds) A mathematical introduction to string theory, S. ALBEVERIO, J. JOST, S. PAYCHA, S. SCARLATTI Novikov conjectures, index theorems and rigidity I, S. FERRY, A. RANICKI & J. ROSENBERG (eds) Novikov conjectures, index theorems and rigidity II, S. FERRY, A. RANICKI & J. ROSENBERG (eds) Ergodic theory of Zd actions, M. POLLICOTT & K. SCHMIDT (eds) Ergodicity for infinite dimensional systems, G. DA PRATO & J. ZABCZYK Prolegomena to a middlebrow arithmetic of curves of genus 2, J.W.S. CASSELS & E.V. FLYNN Semigroup theory and its applications, K.H. HOFMANN & M.W. MISLOVE (eds) The descriptive set theory of Polish group actions, H. BECKER & A.S. KECHRIS Finite fields and applications, S. COHEN & H. NIEDERREITER (eds) Introduction to subfactors, V. JONES & V.S. SUNDER Number theory 1993-94, S. DAVID (ed) The James forest, H. FETTER & B. GAMBOA DE BUEN Sieve methods, exponential sums, and their applications in number theory, G.R.H. GREAVES, G. HARMAN & M.N. HUXLEY (eds) Representation theory and algebraic geometry, A. MARTSINKOVSKY & G. TODOROV (eds) Clifford algebras and spinors, P. LOUNESTO Stable groups, FRANK O. WAGNER Surveys in combinatorics, 1997, R.A. BAILEY (ed) Geometric Galois actions I, L. SCHNEPS & P. LOCHAK (eds) Geometric Galois actions II, L. SCHNEPS & P. LOCHAK (eds) Model theory of groups and automorphism groups, D. EVANS (ed) Geometry, combinatorial designs and related structures, J. HIRSCHFELD et al Number theory 1992-93,
London Mathematical Society Lecture Note Series. 241
Surveys in Combinatorics, 1997 Edited by
R. A. Bailey Queen Mary and Westfield College, University of London
AMBRIDGE
UNIVERSITY PRESS
CAMBRIDGE UNIVERSITY PRESS Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo
Cambridge University Press The Edinburgh Building, Cambridge CB2 2RU, UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521598408
© Cambridge University Press 1997
This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 1997
A catalogue record for this publication is available from the British Library ISBN-13 978-0-521-59840-8 paperback ISBN-10 0-521-59840-0 paperback Transferred to digital printing 2005
Contents M13 by J. H. Conway 1 Introduction . . 2
3
4 5
6 7 8
1 .
.
Definition of M13 . The first few moves
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
The hexads . . . . . The doublings of M13 The inner product . The proofs . . . . . . Further comments . .
References
.
.
.
.
.
.
.
.
1
.
.
.. .. ..
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
.
.
.. ..
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 8 9
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
10
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
..
.
..
11
The Harmonious Chromatic Number and the Achromatic Number by Keith Edwards 1
2 3
4 5
6 7 8 9
Introduction . . Basic properties General bounds
.
.
.
. .
13
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Special graphs . . . Asymptotic results
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
23
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Trees ................................. 26 .
.
.
.
Computational complexity Related topics . . . . . . . Open problems . . . . . .
References
.
.
.
.
.
.
.
.
.
.
.
.
.. .
.
..
Computer Construction of Block Designs by Clement Lam 1
2
Introduction . . . . . . . . . . . . Formulating the problem for BDX . .
Towards a search Conclusion . . . . References . . . . . . . 3 4
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.. .. .. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.. .. .. . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
3 4
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . Partitions and quotients of permutation groups and graphs Primitive graph quotients . . . . . . . . . . . . . . . . . .
.
.. .. ..
..
.
.
49 50 56 63 63
65 .
. .
.
.
.
.
.
.
Normal quotients of locally primitive and locally quasiprimitive . . . . . . . . . . . . . . . . . . graphs . . . . . . . . . . Finite quasiprimitive permutation groups . . . . . . . . . . . . 5 Finite quasiprimitive 2-arc transitive graphs . . . . . . . . . . 6 Full automorphism groups of quasiprimitive graphs . . . . . . 7 . . . . . . . . . . . . . . . . . . . . . . . . . . References . . .
..
34 38 40
49
.
Finite Quasiprimitive Graphs by Cheryl E. Praeger 1
13 15 17 20
. . .
. .
65 65 67 70 72 76 79 82
Contents
vi
Tree Width and Tangles: A New Connectivity Measure and Some Applications by B. A. Reed 87 1
2
3
4 5
6 7
Introduction . . . . . . . . . Graphs of bounded tree width Untangling tangles . . . . . Excluding walls . . . . . . . Graph minors revisited . . . Packing and covering . . Building a wall . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
References
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.. ..
87 . 105 . 119 . 128 . 130 . 136 . 145 . 158 .
Minor-monotone Graph Invariants by Alexander Schrijver
163
1
Introduction .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
Overview of µ(G)
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
Overview of A (G)
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
Some basic facts on µ(G) . . . . . . . . . . . . µ(G) and A(G) for complete graphs . . . . . . . Clique sums . . . . . . . . . . . . . . . . . . . . Behaviour of µ(G) and A (G) under YA and AY µ(G) and A(G) for complete bipartite graphs . . Characterizing A(G) < 1 and µ(G) < 1 . . . . . Van der Hoist's lemma . . . . Characterizing ii(G) < 2 and A(G) < 2 . . . . . Characterizing p(G) < 3 . . . . . . . . . . . . Characterizing .\(G) < 3 . . . . . . . . . . . . A Borsuk theorem for antipodal links . . . . . Characterizing µ(G) < 4 . . . . . . . . . . . . . Towards characterizing A(G) < 4 . . . . . . . . An extension to oriented matroids . . . . . . . . The related graph invariant ic(G) . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 163 . 164 . 166 . 169 . 175 . 176
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
. .
.
.
5
6
7 8 9 10 11
12 13 14 15 16 17 18
.
.
.. .. .. .. ..
.
.
.
References
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.. ..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
178 178 179
. 179 . 181 . 181 . 182 . 186 . 187 . 189 . 191 . 192 . 194
Some Applications of Algebraic Curves in Finite Geometry and 197 Combinatorics by T. Szonyi 1
2
3
4 5
6
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 . . . . . 199 Weil's theorem and its variants . . . . . . . . . . . Applications of Weil's theorem in graph theory and combinatorics201 Applications of Weil's theorem in finite geometry . . . . . . . . 207 The generalized Menelaus' theorem and applications: large arcs 215 The Redei polynomial and applications: blocking sets and (k, n)-
..
arcs .................................220
References
.
.
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 228
Contents
vii
New Perspectives on Interval Orders and Interval Graphs by William T. Trotter 1
2 3
4 5 6
7 8 9 10 11
12 13 14 15 16 17 18 19
237
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . Interval orders and interval graphs . . . . . . . . . . . . . . . Classical representation theorems . . . . . . . . . . . . . . . . Dilworth's theorem for interval orders . . . . . . . . . . . . . . Linear extensions and dimension . . . . . . . . . . . . . . . . . Linear extensions of interval orders . . . . . . . . . . . . . . . Dimension of interval orders . . . . . . . . . . . . . . . . . Critical pairs and alternating cycles . . . . . . . . . . . . . . . Interval orders and shift graphs . . . . . . . . . . . . . . . . . Interval orders and overlap graphs . . . . . . . . . . . . . . . . Semi-orders and balancing pairs . . . . . . . . . . . . . . . . . Interval orders and extremal problems . . . . . . . . . . . . . Interval orders and hamiltonian paths . . . . . . . . . . . . . . On-line and un-cooperative coloring . . . . . . . . . . . . . . . Fractional dimension and ramsey theory for probability spaces Higher-dimensional analogues for graphs . . . . . . . . . . . . Higher dimensional analogues for orders . . . . . . . . . . . . Intervals, angles and spheres . . . . . . . . . . . . . . . . . . Tolerances, thresholds and gaps . . . . . . . . . . . . . . . . .
References
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
.
.
Approximate Counting by Dominic Welsh Introduction . . . . . . . . . . . . 2 Counting to within a ratio . . . . Randomised approximation schemes 3 4 Rapidly mixing Markov chains . . . Computing in a convex body . . . . 5 Partial orders . . . . . . . . . . . 6 Graph problems . . . . . . . . 7 Contingency tables . . . . . . . . 8 Matroid problems . . . . . . . . . . 9 10 Zonotopes . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . 1
..
. 237 . 239 . 239 . 243 . 244 . 245 . 246 . 248 .
249
. 252 . 255 . 257 . 259 . 261 . 264 . 266 . 269 . 271 . 272 . 277
287
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
. 287 . 289 . 291 . 294 . 297 . 300 . 302 . 308 . 309 . 313 . 317
Author Index
325
Subject Index
331
ix
Preface The 1997 issue of the British Combinatorial Bulletin contains a short history, written by Norman Biggs, of the early years of the British Combinatorial Conference. The first one was held at Oxford in 1969. The sixth conference, held at Royal Holloway College in 1977, was the first at which a volume containing the invited talks was published in time to be available to participants at the conference. Peter Cameron was the pioneering editor of that volume. Such a volume has been produced for every conference thereafter. The 1977 conference was also the first one that I attended. There I joined the British Combinatorial Committee, which was formally set up at that meeting although it had effectively existed for some years-the previous conferences didn't just organize themselves. As often happens, I found that being on the committee considerably widened my knowledge of the subject. I left the committee in 1981, but have never lost touch with combinatorial activity in Britain. I was delighted when I was asked to edit the present volume. In spite of the work involved, I am still delighted. I have had a preview of nine magnificent
papers, and come to know their subject matter much better than I would otherwise have done.
At the centre of this volume is a long paper by Bruce Reed about the tree width of graphs. This is a new measure of connectivity. It is intimately linked to the concept of a minor of a graph, which is obtained by erasing an edge or coalescing two vertices joined by an edge, or by a sequence of such operations. Although the idea of `forbidden minors' was made famous by Kuratowski's characterization of planar graphs in 1930, the main theoretical development of graph minors has taken place over the last decade, led by Neil Robertson and Paul Seymour. At the 1985 British Combinatorial Conference in Glasgow, Seymour talked about the early stages of this work, hot off the press. Now
Reed, who has himself been one of the contributors to the area, gives this splendid survey of exceptionally deep, interesting and valuable work, some of the most important work ever in graph theory. It shows that tree width and
minors give a rich mathematical theory to graphs in a way that the more obvious concepts of k-connectivity and induced subgraphs do not. It is a long
paper for a conference proceedings, but it is a magnificent exposition of a major piece of work. It is laid out in such a way that the non-specialist can read it with ease, while at the same time containing proofs of the important results. This should become the definitive paper on the topic. Alexander Schrijver's paper is a natural accompaniment to Reed's, because it too is concerned in part with graph minors. He describes some new and intriguing graph parameters which are non-decreasing upon taking minors. One of these, introduced by Colin de Verdiere in 1990, gives another characterization of planarity but also has wider applications. A related parameter was introduced by Schrijver and co-workers in 1995. Colin de Verdiere's invari-
x
Preface
ant was suggested by ideas from differential geometry;, most of the related invariants are more obviously combinatorial. Schrijver gives simpler proofs than those in the literature and poses interesting open questions about the relationship between these parameters (and others). Perhaps the most famous graph parameters are the chromatic number and chromatic index, the smallest number of colours with which the graph can be properly vertex-coloured and edge-coloured respectively. In a proper vertexcolouring, the colours on the the ends of an edge must be different. Natural restrictions on such a colouring are to demand that each unordered pair of colours is used on at most one edge or at least one edge. These give harmonious and complete colourings respectively. Keith Edwards gives a clear and interesting survey of work on the parameters associated with these two types of colouring, some of it extremely recent. The paper will form a very useful background for further research in this area. One use of graphs is as tool to study partially ordered sets (posets). In a poset, for each pair x and y of distinct elements, either x < y or y < x or x and y are incomparable. The comparability graph of the poset is obtained by joining x to y whenever x and y are comparable. Some properties of posets are constant over all posets with the same comparability graph-comparability invariant. In applications such as scheduling of tasks or dating archaeologicial finds, the basic objects are intervals on the real line. These can be turned into posets by considering when the whole of one interval is to the left of the whole of another. Thus begins the theory of interval orders. Or they can be turned into graphs-interval graphs-by joining any two intervals whose intersection is not empty. Tom Trotter has contributed extensively to the theory of posets and its links with graph theory. Here he gives an excellent survey of the topics of interval orders and interval graphs. Although it is also about graphs, Cheryl Praeger's paper has quite a different focus. Her motivation is groups of automorphisms of graphs. She considers highly symmetric graphs, those which are at least vertex-transitive (i.e. which admit groups of automorphisms which are transitive on the vertices.) Often
she demands that the graph be 2-arc transitive, which means that its automorphism group is transitive on ordered paths of length two. A well-studied stronger property than transitivity is primitivity: a group acts primitively if it preserves no non-trivial partition. Praeger defines a quasiprimitive group action as a natural generalization of a primitive group action, and defines a graph to be quasiprimitive if it admits a quasiprimitive group of automorphisms. Surveying work in which she and her collaborators have played a major role, she gives a complete categorization of quasiprimitive groups into eight types, parallel to the categorization given for primitive groups by the celebrated O'Nan-Scott theorem. She summarizes what has been done so far towards the classification of quasiprimitive graphs, especially those that are 2-arc transitive. This is an excellent example of
Preface
xi
how advances in permutation group theory following the classification of finite simple groups have made feasible such classifications of graphs. This is a very clear account of work still in progress. Of course, automorphisms are a natural tool for studying all kinds of combinatorial structures, not just graphs. Sometimes we want to find all structures
admitting an automorphism group with certain properties, as in Praeger's paper. At the other extreme, sometimes we study one particular example of a combinatorial structure precisely because its automorphisms are so interesting. John Conway's paper takes this point of view. It presents a new construction of the Mathieu group M12, one of the simplest available. Most constructions of sporadic simple groups such as M12 either produce some combinatorial structure (Steiner system, Hadamard matrix, code, ...), show that it is unique and deduce properties of its automorphism group (so that the actual automorphisms are not easy to construct), or else produce the
permutations directly in such a way that it is hard to see the combinatorial structure. Conway's new construction takes the latter approach-all the permutations are produced at the outset-but the combinatorial properties are surprisingly easy to verify. Indeed, the exposition is skilfully constructed: once the descriptive parts of the paper have been read, the proofs are almost obvious.
A byproduct of this approach is the result that there exists a sharply 6transitive set of permutations of 13 objects. They are the permutations of 12 counters and one hole in an analogue of the 15-puzzle played on the 13-point
projective plane. This result will be interesting to researchers on extremal combinatorics, specifically, metric properties of sets of permutations. Clement Lam's use of automorphisms is more conventional. The setting is 2-designs, also known as as balanced incompete-block designs, or just block designs-collections of k-subsets (called blocks) of a v-set with the property that every 2-subset is contained in the same number of blocks. Here is a very
useful article about the use of the BDX program (demonstrated at the conference) to search for block designs. The reader is taken through an extended worked example-the search for designs which have a certain set of parameters and which admit an automorphism of order seven with no fixed points or blocks-and sees details of calculations which are not usually provided in published articles on mathematics. The reader is led so gently through the example that (s)he should have no difficulty in trying out BDX alone. One of the classical sources of block designs is projective geometry: the points are the points of the geometry and the blocks are subspaces or conics, possibly with additional points. Thus we are led to questions in geometry such
as: how may points are in the intersection of an object of one sort with an object of another sort? What size can a minimal blocking set be? (A blocking
set is a set of points which meets every line.) Tams Szonyi's paper is a wide-ranging and detailed survey of such questions. It shows many important applications of two important results-Weil's theorem and Segre's lemma of
xii
Preface
tangents. It updates the material on blocking sets reported by Aart Blokhuis at the 1993 British Combinatorial Conference at Keele. A variety of results is described. The flavour of the proofs is given, rather than their technicalities, and further probable developments are discussed. The papers by Edwards, Reed, Schrijver and Trotter all refer to problems in computational complexity. A typical problem is the following: given a graph G with n vertices, can we decide whether G is Hamiltonian? More specifically, is there an algorithm that will decide the answer within time that is linear in n, or polynomial in n, or otherwise bounded by some function of n? During the past decade or so a new type of complexity problem has been considered: not "can we?" but "how many?" and not "exactly" but "approximately". A typical example now is: given a graph G with n vertices, how many forests does G contain as induced subgraphs? Here "how many" means to within an order of magnitude. Once again one wants to know. what sort of function of n bounds the running time for an algorithm that answers the question. Dominic Welsh's paper provides a fascinating survey of the very interesting material in this area, and thus rounds off this volume of papers. In 1977 the task of the editor was to obtain typescripts from the speakers, by a deadline, and to submit camera-ready copy to the publisher. Over the twenty intervening years, what has changed? The deadlines are still there, and so is a publisher, even though Cambridge University Press has replaced Academic Press. Typescripts, with mathematical symbols either hand-written or typed from 'golf-balls', have vanished, being replaced first by word-processors then by mathematical type-setting packages such as TjX. As this happened, camera-ready copy at first became more diverse then converged as mathematicians came to agree on the use of a few systems. These proceedings are, I believe, the first in this series for which the editor has made a serious attempt at uniformity by asking all the authors to use the same system and giving them a style file to encourage them. The authors have cooperated marvellously, often to tight schedules. Although they had not all used LATEX before, they all submitted their papers in fairly standard I?TJY, using a rather simple style file which I had provided. My heartfelt thanks to them all. None of this would have been possible without 1 EX, specifically the current version of WT, which is designed for people like me who are fussy about what
their written mathematics looks like but do not want to program anything much more complicated than `this is the end of the statement of the theorem'. So thanks to Donald Knuth for giving TjX to the world; to Leslie Lamport, for the original version of YlE and for the manual which taught me to think in terms of generic mark-up; and to the I?TEX3 team, led by Frank Mittelbach and Chris Rowley, for upgrading and maintaining WI , for incorporating features that I asked for, and for answering lots of my questions. The other big change over the twenty years is the arrival of electronic mail.
All the papers were submitted by email. Almost all of them were sent on
Preface
xiii
to referees by email, and reports came back by the same route. During the editing phase the myriad queries such as `Do you mean \log or \ln there?' were dispatched and answered promptly, all by email. Without email it would simply have been impossible to do the task within the given time-scale. I am grateful to the authors and the referees (as well as various people who helped me with the names of journals and books unfamiliar to me) for dealing with everything as quickly as they did. In editing these proceedings I have tried to strike a fine balance between respecting the authors' wishes about how their mathematics is presented and providing the uniformity of layout that, by melting into the background, enables the reader to concentrate on the content and so read it more easily. If I have done my job well its effects should be invisible. Finally, thanks to numerous people in the School of Mathematical Sciences at Queen Mary and Westfield College. Not only have they provided hardware,
software, advice and moral support. They have been very patient with me while other jobs have been left on one side as I worked on these proceedings. This is particularly true of the other members of the local organizing team of this British Combinatorial Conference-Peter Cameron, Leonard Soicher and Shirley Wilkinson. Rosemary A. Bailey School of Mathematical Sciences Queen Mary and Westfield College Mile End Road London El 4NS
[email protected] 16 April 1997
M13 J. H. Conway Summary The group M12 has no transitive extension, but the object of the title is the next best thing: a set of permutations which is an extension of M12. We give an elementary construction, based on a moving-counter puzzle on the projective plane of order 3, and provide easy proofs of some of its properties. 1
Introduction
Long ago I was intrigued by the fact that M12, E. Mathieu's celebrated quintuply transitive group on 12 letters, shares some structure with L3(3), which acts doubly transitively on the 13 points of the projective plane PG(2, 3), of which it is the automorphism group.
To be more precise, the point-stabilizer in L3(3) is a group of structure 32: 2S4 that permutes the 12 remaining points imprimitively in four blocks of 3, and there is an isomorphic subgroup of M12 that permutes the 12 letters in precisely the same fashion. Again, the line-stabilizer in L3(3) is a group of this same structure that permutes the 9 points not on that line in a doubly transitive manner, while the stabilizer of a triple in M12 is an isomorphic group
that permutes the 9 letters not in that triple in just the same manner. In the heady days when new simple groups were being discovered right and left, this common structure inevitably suggested that there should be a new group that contained both M12 and L3(3), various copies of which would
intersect in the subgroups mentioned above. Of course this turned out not to be the case, but some years ago I found an almost equally satisfactory explanation - M12 and L3(3) are indeed both subgroups of the same object, but that object is not a group! I call it M13. Sections 2-6 will be purely descriptive, and contain numbered assertions, which will be proved in Section 7. 2
Definition of M13 Since
M12 is a set of 95040 = 12.11.10.9.8 permutations of 12 letters,
(1)
we might expect that
M13 is a set of 1235520 = 13.12.11.10.9.8 permutations of
(2)
13 letters;
and since M12 is quintuply transitive,
(3) 1
J. H. Conway
2
we might hope that M13 is sextuply transitive.
(4)
These expectations turn out to be true, but we must be careful about the meanings of the terms.
Unfortunately, the word "permutation" retains two distinct senses - it may refer either to a particular arrangement of n objects, or to a particular operation of rearranging them, the latter usage being common among group theorists and the former among the public at large. We shall define M13 to be a particular set of permutations of 13 objects in the lay sense, namely certain ways of putting 12 lettered counters and one hole on the points of a projective plane P = PG(2, 3). We can think of this in terms of a "13-puzzle" analogous to Sam Loyd's famous 15-puzzle, wherein 15 square tiles and one hole are arranged in a 4 x 4 tray, the object being to proceed from one given arrangement to another by a sequence of moves in each of which the hole is exchanged with one of the adjacent tiles.
In our 13-puzzle, any counter a determines at any time a line that joins its present position to that of the hole. A move of the puzzle is to put the counter into the hole (that is, onto the "holy point", as we shall call it), and at the same time to interchange the positions of the other two counters b and c on this line. We shall refer to this as the move albc. To avoid circumlocution, we refer to the point occupied by the counter labelled a as "point a", and the holy point as "point o". M13 consists of all the arrangements (of counters and hole) that can be obtained from a given one by moves of this type. In the language of category theory, M13 is a groupoid, (a category in which all arrows are invertible), whose objects are the 13 positions of the hole, and whose arrows are the rearrangements produced by legal sequences of moves; the initial and terminal objects of an arrow are the positions of the hole before and after the moves are made. 3
The first few moves
Figures 1-4 show the way we shall draw the projective plane P. In Figure 1, the points of the plane are numbered 0-11 and oo, and to avoid awkward bends some of the points on some lines are indicated by hooks. We have chosen this particular way of drawing the plane so as to emphasize the close relationship with the "MINIMOG" array (see Figure 2). Figure 3 shows the usual coordinatization of the plane. The 9 points on the right are those of the corresponding "Euclidean" or "affine" plane, the point marked XY being that with affine coordinates (X, Y) or projective coordinates
(X, Y, 1). This is extended to the projective plane by adjoining the "line at
Figure 1: PG(2, 3)
6
3
0
9
5
2
7
10
4
1
8
11
Figure 2: The MINIMOG
infinity" on the left, whose point marked m (for m = 0, 1, 2) has projective coordinates (1, m, 0) and lies on the three parallel lines y = mx + c, while that marked oo has coordinates (0, 1, 0) and lies on the three "vertical" lines x = c. In Figure 4 we display only the y-coordinates of the points other than oo. There are 4 "vertical" lines (passing through the point oo), and the 9 others meet these in the points determined by one of the 9 words 0 000, 0 111, 0 222, 1 012, 1 120, 1 201, 2 021, 2 102, 2 210
of the "tetracode" (see [3], [2]). The typical tetracode word is
m cc+mc+2m
00
00
Figure 3: Coordinates
Figure 4: Tetracode coordinates
J. H. Conway
4
in which the last three digits form an arithmetic progression (mod 3), whose "slope" m is the first digit. For example, the line {5, 3, 7,11} of Figure 1 is called 1 012 in Figure 4.
It should be obvious that the set of permutations of the counters that can be achieved by move sequences that restore the hole to its original position at oo forms a group; anticipating a later result, we call this "the group M12". Figure 5 shows a few successive moves in the 13-puzzle, whose effect is
to move the hole around a triangle and restore it to its original position, interchanging the four pairs (4, 5), (3, 0), (6, 9), (10,11)
of counters as it does so. This gives us an element of M12. We call a permutation of this sort a triangular permutation.
4
5
645
V/1 6
903
6
611011-+
(4 5) (6 9) (0 3)(10 11) E M12
Figure 5: A triangular permutation Look at the action of this permutation on the middle two of the four "vertical" lines. The points 1, 2, 7, 8 are fixed, and 3 and 0 are interchanged. By symmetry, we see that M12 contains a permutation that swaps any two points lying on distinct verticals, and leaves the other four points on these verticals unchanged. This shows that M12 acts transitively on the 12 counters, since we can use one of the above moves to take any given counter off the leftmost vertical (if
necessary), and then another to bring it back to the topmost point of that vertical. Indeed it also establishes double transitivity, since two more such moves (at most) are needed to bring the second of two given counters to the second point of the leftwise vertical. One could prove triple transitivity in the same way, but this will not be necessary. 4
The hexads
The classical M12 is known to permute a collection of 132 hexads that form a so-called Steiner system S(5, 6,12). How do we recognize these in our picture?
5
M13
4.3.3=36
12. (2) =36
(2)'32=54
Figure 6: Hexads Figure 6 shows the answer. At any instant, two lines contain either 6 counters and the hole, in which case these counters form a hexad; or 7 counters, from which we form a hexad by removing the counter at the intersection of the two lines. There is a further type of hexad, consisting of the points (other than the vertices) lying on the lines of a triangle one of whose vertices is the holy point. The Figure shows the four different geometrical appearances that a hexad can assume, and counts the hexads, showing that there are 132 in all. Note that the triangular permutation shown in Figure 5 fixes the hexad {0, 1, 2, 3, 7, 8} and induces the transposition (0 3) on it. One could easily give a case-by-case verification that.
a move of the 13-puzzle takes every hexad to another hexad, although possibly of a different shape.
(5)
One could also check that
no two distinct hexads can contain the same 5 points, and hence the hexads form an S(5,6,12),
(6)
and that M12 acts transitively on the hexads. 5
(7)
The doublings of M13
The group M12 is known to have Schur multiplier of order 2. This reveals itself by the existence of a group 2M12 that has a homomorphism onto M12 with kernel of order 2. This group can be constructed as a group of monomial permutations of ±12 letters: that is to say, it permutes 24 symbols, say
+a, -a, +b, -b, ... , +k, -k, +1, -l in such a way that we obtain the permutations of M12 simply by ignoring the signs. The non-trivial element of the kernel negates all 24 symbols. Is there an analogous 2M13?
J. H. Conway
6
Indeed there is! We can obtain it by labelling the opposite sides of each of our counters with the two signed versions of the appropriate letter. But now when we move some counter into the hole, the two other counters that we interchange must also be turned over. We call this the ±13-puzzle.
It turns out that the monomial permutations realized by sequences of the new moves that return the hole to its original position oo do indeed form the usual group 2M12.
(8)
Moreover,
2M12 is doubly transitive in the monomial sense; that is, given two pairs a, b and c, d of distinct letters and any signs a, /3, y, 6, there is an element of 2M12 which maps as to yc and /3b to 8d.
(9)
Since we have already shown the double transitivity, it is enough to show that 2M12 contains an element fixing one counter of the ±13-puzzle and reversing another. Such an element is easily discovered. For example, the product of the triangular permutations obtained from the triangles oo69, oo30, and oo6 10 fixes 4 and negates 3.
The automorphism group of the classical M12 is a group M12.2 that permutes 24 letters in two sets of 12, with the properties each permutation of the original M12 extends uniquely to a per-
(10)
mutation of the 24 letters, and an outer automorphism interchanges the two sets of 12 letters. Is there an analogous M13.2? Indeed there is! To obtain it, we enlarge the 13-puzzle to the "26-puzzle" by adjoining an additional hole Q and 12 new counters
A,B,...,K,L that we associate with the lines of P in the same way as the old ones are associated with the points. However, we also demand that the "holy point" must always be incident with the "holy line". This entails that whenever we make a point-move albc, the four points involved must form the holy line. Dually, we now have "line-moves", say ABC, for which the four lines involved must pass through the holy point. Any sequence of moves in the 26-puzzle yield a legal sequence of moves in the 13-puzzle if we just ignore the line-counters. It might seem that the incidence condition would restrict our freedom; but in fact an arbitrary sequence of point-moves can still be performed, by interleaving them with the appropriate line-moves. Thus, if we wish to follow a point move on the line L1 by another
7
M13
on a different line L2, we need merely interpose the line-move that moves the hole from L1 to L2. In this way, for any permutation of M12 we can find a sequence of alternating point and line moves that effects that permutation of the point counters, and also restores both holes to their original positions (say, the point oo and
the line at infinity). It turns out that the resulting permutation of the line-counters is uniquely determined, giving the extension of M12 to a group on 24 letters, doubly transitive on both sets of counters.
(11)
We therefore define M13.2 to be the set of all permutations of the 26 counters and holes obtainable from the starting position by moves of the 26-puzzle. The group M12.2 has a double cover 2M12.2 containing the group 2M12 that we described earlier. This also has an analogue, 2M13.2, obtained by using both
sides of both sets of counters. Provided that o, a, b, c lie on the current holy line Q, we can make a point-move albc that puts a in the hole and interchanges and negates both b and c; dually, provided Q, A, B, C pass through the current holy point o, we can make the line-move AJBC that puts A in the hole and interchanges and negates B and C. We call it the "±26-puzzle".
= point, Figure 7: The incidence graph of PG(2, 3)
* = line
J. H. Conway
8
6
The inner product We regard the symbols on the counters of the ±26-puzzle as vectors ±u (for
points) and ±V (for lines), and introduce an inner product between the two sets. We define (u, V) to be -1 just if either u and V are incident with each other, or both are incident with the current holes; otherwise (u, V) = +1. In terms of the incidence graph of the plane (which has vertices that correspond to the points and lines, with edges corresponding to the incident pairs, see Figure 7), (u, V) is equal to -1 just if either is accessible from the other that is to say, if there is a path from u to V that contains no other counter. It is quite remarkable that this inner product is unchanged by making any legal move of the ±26-puzzle.
(12)
To see this, look at the effect of a point-move albc in the incidence graph. Figure 8 shows that the same line-counters N, P, Q, R, S, T are accessible from a
before and after the move, while those accessible from either b or c before the move are precisely those not accessible from -b or -c after it. (Thus, R, S, T, U, V, W are initially accessible from b, while N, P, Q, X, Y, Z become accessible from -b.)
It turns out that the hexads that were introduced in an ad hoc way earlier can now be given a simple uniform definition.
(13)
If V and W are labels from any two line-counters, there are 6 point-counters whose labels u satisfy (u, V) = (u, W), which form a hexad [V, W]; the remaining 6 satisfy (u, V) = -(u,W) and form the hexad [V, -W] = [-V, W]. Let H be the Gram matrix of the inner product: its rows and columns are indexed by points and lines, the entry in row u and column V being (u, V).
U*
U*
R
N
41-c
*W
R S
Xcz Y
T
0
T
X
b Z
Y
Figure 8: Invariance of the inner product
9
M13
Then any two columns of H agree and disagree in sets of points forming hexads.
So HT H = 121, and H is a Hadamard matrix. 7
The proofs
It is trivial to check (13): in the new notation, the representative hexads of Figure 6 are as shown in Figure 9. Again, since the hexads are defined in terms of the inner product, (12) immediately implies (5). We can also see that they must be permuted transitively (as claimed in (7)): to move [V, W] to [X, ±Y] it suffices to move V to X and W to ±Y (using the dual of (9)).
[X, Y]
[V, -W]
[V, -W]
Figure 9: Hexads defined by the inner product
We now prove (10) and (11). Let P and Q be the matrices representing the monomial permutations of points and lines derived from any sequence of
moves. Then P-'HQ = H, by (12). In particular, Q = H-1PH is uniquely determined by P (since H is invertible). The interchange of point and line counters is effected by a polarity of the plane which swaps the positions of the two holes.
Consider a sequence of point and line moves which leaves every point counter fixed (possibly reversed). Then P is a diagonal matrix with p2 = I. Suppose that P 0 ±I. Since HTPH = 12Q, we see that P has six entries +1 and six entries -1, and the set of six positions where the +1s occur meets every hexad in 0, 3 or 6 points. However, no such set can exist. So P = Q = ±1. The proof of (6) is most easily seen using the Hadamard matrix H, or (for
convenience of notation) HT. By column sign changes, the first three rows of HT may be assumed to be
+ + + + + + + + + + + + + + + + + + - - - - - + + + - - - + + + - - If these rows are V, W, X, then clearly [V, ±W] fl [V, ±X] = 3. For any other
row Y, if Y has a entries + among the first three, then [V, ±W] fl [X, ±Y]
J. H. Conway
10
is 2a or 6 - 2a, which is even in either case. So no two hexads can meet in five points. The average number of hexads containing a set of five points is 132.6/ (5) = 1; so any five points lie in a unique hexad. What is the order of M12? We know by (7) that M12 acts transitively on hexads. Moreover, the transpositions induced on the hexad {0, 1, 2, 3, 7, 8} by
triangular permutations as in Figure 5 generate S6. Further, the group K fixing a hexad pointwise is trivial (see below). So IM121 =132.6!=12.11.10.9.8.
Thus, (1) holds. The quintuple transitivity of M12 follows: give two 5-tuples, we find g E M12 mapping the hexad R containing the first tuple to the hexad 9-l' containing the second; then we may use the symmetric group on W to move the tuple to its required position. Thus, (3) holds. The corresponding facts (2) and (4) about M13 hold because M13 is the union of 13 translates of M12, one for each possible position of the hole. The fact that the pointwise stabilizer K of a hexad 7d is trivial follows from
the structure of the Steiner system. Given a dead {x, y} C f, the three sets f' \ W, for hexads 9d' D 7l \ {x, y}, form a syntheme on the complement of 9i (a partition into three sets of size 2). Distinct duads give distinct synthemes. Now K fixes all duads on 1-l, and hence all synthemes on the complement; so
K = 1. Finally, the identification of M12 and the related groups with the "classical" versions follow from properties of the classical versions: for example, there is a unique Steiner system, whose automorphism group is M12; and there is a unique Hadamard matrix whose group of monomial automorphisms is 2M12 (see [2]).
8
Further comments In view of the fact that 13.132 = 1716 =
0131 6
'
we might wonder if it is possible to partition all the 6-element subsets of a set of size 13 into disjoint copies of S(5, 6,12), each omitting one point of the set. In fact this is impossible, even though the analogous partition of 4-element subsets of a 9-element set can be done in two distinct ways [1]. So it is worth seeing how close we get. For each position of the hole, the 132 hexads define a S(5, 6,12) on the remaining points. However, not every 6-element set occurs as a hexad in one of these Steiner systems, and some occur more than once. Ignoring the hole, the hexads are of three geometric types. The union of two lines with the intersection removed occurs as a hexad for 7 positions of the hole; the three sides of a triangle with the vertices removed, for 3 positions; and the union of two lines with a point on one line removed, just once.
M13
11
Similar puzzles can be defined on some other configurations, and also define
interesting groups. If we use the projective plane of order 2, so that a move consists of a cyclic permutation of the hole and two counters on a line with it, we obtain the alternating group A6. We can realise its triple cover 3A6 by using 3-sided counters (rotated in opposite directions during a move), and a group A6.2 by using line-counters as well as point-counters. On the Petersen graph, we can play the game using two holes, which must always be adjacent. A move consists of permuting cyclically the three neighbours of a hole (the other hole and two counters). We obtain the group L2(7) (if the holes return to their original positions) or L2(7).2 = PGL(2, 7) (if they are allowed to be interchanged by the move sequence).
References [1] P. J. Cameron & C. E. Praeger, Partitioning into Steiner systems, in Com-
binatorics '88, Volume 1 (eds. A. Barlotti et al.), Mediterranean Press, Rende (1991), 61-71. [2]
J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker & R. A. Wilson, An ATLAS of Finite Groups, Oxford University Press, Oxford (1985).
[3] R. T. Curtis, The Steiner system S(5, 6,12), the Mathieu group M12, and the "Kitten", in Computational Group Theory (ed. M. Atkinson), Academic Press, (1984).
Department of Mathematics Princeton University Princeton, NJ 08544 U.S.A.
The Harmonious Chromatic Number and the Achromatic Number Keith Edwards Summary The harmonious chromatic number of a graph is the least number of colours in a vertex colouring such that each pair of colours appears on at most one edge. The achromatic number of a graph is the greatest number of colours in a vertex colouring such that each pair of colours appears on at least one edge. This paper is a survey of what is known about these two parameters, in particular we look at upper and lower bounds, special classes of graphs and complexity issues. 1
Introduction
A short survey of harmonious colourings was given by Wilson [80] in 1990. Since then a number of new results have appeared, and the close relationship between harmonious chromatic number and achromatic number has been observed. The purpose of this new survey is to outline what is known about these parameters, and suggest some open problems. A more detailed summary of results on the achromatic number, with a rather different emphasis, can be found in the forthcoming survey by Hughes and MacGillivray [51]. We begin with the definitions of the two parameters.
Definitions A harmonious colouring of a graph G is a proper vertex colouring of G such that, for any pair of colours, there is at most one edge of G whose endpoints are coloured with this pair of colours. The harmonious chromatic number of G, denoted h(G), is the least number of colours in a harmonious colouring of G.
Definitions A complete colouring of a graph G is a proper vertex colouring of G such that, for any pair of colours, there is at least one edge of G whose endpoints are coloured with this pair of colours. The achromatic number of G, denoted O(G), is the greatest number of colours in a complete colouring of G.
Remark Although these colourings have usually been called complete colourings, the terms irreducible colouring [64], full colouring [32, 46], correct colouring [8] and achromatic colouring [84] have also been used.
The achromatic number was introduced by Harary, Hedetniemi and Prins [41] in 1967. They considered homomorphisms from a graph G onto a complete
graph Kn. A homomorphism from a graph G to a graph G' is a function ¢: V (G) -* V (G') satisfying u N v = uq - vo. This induces an obvious mapping from E(G) to E(G'). It is easy to see that a complete colouring of G with n colours corresponds precisely to a complete homomorphism of G 13
14
Keith Edwards
onto K,z, i.e. one whose induced edge mapping maps E(G) onto E(Kn,). They noted that the smallest n for which such a complete homomorphism exists is just the chromatic number x(G) of G. They considered the largest n for which such a homomorphism exists. This was later named the achromatic number '(G) by Harary and Hedetniemi [40]. In the first paper [41] it is shown that there is a complete homomorphism from G onto Kn if and only if n satisfies
x(G) < n < %(G). In a later paper, Frank, Harary and Plantholt [33] introduced the concept of a line-distinguishing colouring. This is similar to a harmonious colouring, for any pair i, j of distinct colours, there is at most one edge with endpoints coloured i, j, but in addition, for each colour i, there may be at most one edge with both endpoints coloured i. Thus a line-distinguishing colouring need not be proper. The line-distinguishing chromatic number A(G) is defined in the obvious way as the least number of colours in such a colouring of G. The authors attribute the definition to Pierre Duchet. It was shown by Hopcroft and Krishnamoorthy [48] that determining A(G) is NP-hard, and a short proof of the same result by David S. Johnson is also given in the paper. They refer to A as the harmonious chromatic number, but this name has subsequently been used for the definition which we have given above. Although there is an obvious similarity (or perhaps duality) between the definitions of h and 0 given above, this does not appear to have been noticed until recently. There are several other ways of defining the harmonious chromatic number. A harmonious colouring of G with n colours corresponds to a homomorphism from G to Kn such that the induced map from E(G) to E(Kn) is one-to-one. Then h(G) is just the least n for which such a map exists. Another way of looking at this is the following: Given a graph G, if we identify two adjacent vertices of G we get a loop, and if we identify two vertices at distance 2, we get a multiple edge. However if we identify two vertices
distance 3 or more apart, we get a new simple graph. We can repeat the process until the resulting graph has diameter 2. Then h(G) is just the smallest number of vertices in a diameter 2 graph which can be obtained from G in this way.
We can also look at the process the other way round. A detachment of a graph G is a graph H which is obtained from G by splitting each vertex into one or more vertices, and such that E(H) = E(G) (the edges incident with vertex v in G are "shared out" arbitrarily among the vertices into which v is split). A subdetachment is just a subgraph of a detachment. Then G has a harmonious colouring with n colours if and only if G is a subdetachment of Kn. See Nash-Williams [74] for a survey of detachments.
Notation We will usually denote the number of vertices of a graph by n and the number of edges by m. In particular, m will always mean the number of edges of the graph in question, and this will usually be assumed without
Harmonious Chromatic Number
15
comment. The maximum degree of a graph G is denoted 0(G). 2 2.1
Basic properties Simple bounds
There are two simple but important bounds for h and 0. The first arises from the observation that, in any harmonious colouring of a graph, any vertex,
and all of its neighbours, must all have distinct colours. Hence h(G) > 0(G) + 1. This bound is met for example by a complete graph or a star. More generally, we observe that any two vertices which are adjacent, or have a common neighbour, must have distinct colours. Hence for any graph of diameter 2, the harmonious chromatic number equals the number of vertices, and in general, h(G) is at least the number of vertices in the largest diameter 2 subgraph of G. The other bound arises from a comparison of the number of edges of G and the number of pairs of colours which are available. If G is harmoniously coloured with k colours, then the number of pairs of colours, (2), must exceed the number of edges of G. Similarly, for a complete colouring, the number of pairs of colours must be at most the number of edges. This motivates the following definitions.
Definitions Let m be a positive integer. Then we define Q(m) to be the least positive integer k such that (2) > m, and q(m) to be the greatest integer k such that (2) < m. We also define R(m) to be (QZ"°)) - m and r(m) to be
m - (q(r)). 2 Remark It is easily calculated that Q(m) = 11 + 2m + 11
and
1 + 8m+1
q(m) =
2
1
2
J
Also 0 < R(m) < Q(m) - 1 and 0 < r(m) < q(m). From the discussion above, we have the following crucial result:
Proposition 2.1 For any graph G with m edges, h(G) > Q(m)
and
Vi(G) < q(m).
Much of the interest in these parameters is focussed on attempting to find classes of graphs for which these bounds are attained, or nearly attained. In the special case when m = (2) for some k, there may be a colouring in which each pair of colours appears on precisely one edge. We will call such a colouring an exact colouring. An exact colouring is clearly both a harmonious colouring and a complete colouring. It is clear that for a graph with (2) 2
Keith Edwards
16
edges, the following are equivalent: i) G has an exact colouring, ii) h(G) = k and iii) '(G) = k. This observation, made by N. Cairnie, is the starting point for relating the two parameters. Any result for harmonious colouring involving graphs for which h(G) is near to Q(m) is likely to have a counterpart involving graphs for which '(G) is near to q(m). The correspondence is not however simple in most cases. Achromatic number counterparts of several harmonious colouring results are proved in [19]. 2.2
Partial colourings
Techniques for colouring graphs often involve some sequential colouring of
the vertices, so that at some stage we have a partial colouring of the graph which is subsequently extended to a colouring of the whole graph. For harmonious colourings, it is not the case that a colouring of an induced subgraph can always be extended to a colouring of the whole graph, even if extra colours are allowed. This is because we may have an uncoloured vertex adjacent to two vertices with the same colour. For this reason we make the following definition:
Definition A partial harmonious colouring of a graph G is a colouring of a subset of the vertices of G which is a harmonious colouring of the graph induced by the coloured vertices, and is such that no uncoloured vertex has two or more coloured neighbours with the same colour.
We note that if H is a subgraph of G, then any harmonious colouring of G induces one of H, hence h(H) < h(G). If H is an induced subgraph of G, then a complete colouring of H can always be extended to a complete colouring of G. This can be seen as follows: Suppose that we have a complete colouring of H. Colour the remaining vertices sequentially. Suppose v is the next vertex to be coloured. Then if v has
neighbours with each of the colours used so far, colour v with a completely new colour, otherwise colour v with any existing colour not used on any of its neighbours. In either case the colouring remains complete. We thus have
Proposition 2.2 If H is an induced subgraph of G, then O(H) < z'(G). Note however that we can have a (non-induced) subgraph H of G with
V) (H) > 0(G), for example '(C4) = 2 while 0(P4) = 3. 2.3
Degree sum
Let SS be the set of vertices which have colour c in a harmonious colouring. Then the neighbours of the vertices in SS must all have distinct colours, none of which can be c. Thus the sum of the degrees of the vertices in S, is less than the number of colours. This motivates the following definition:
Harmonious Chromatic Number
17
Definition Suppose that we have a (possibly partial) harmonious colouring of a graph G with C colours, and let S, be the set of vertices with colour c. Then we define the degree sum of colour c to be Eves, d(v). Note that for an exact colouring with C colours, the degree sum of each colour must be exactly C - 1. 3 3.1
General bounds Upper bounds for h(G)
The first upper bound for general graphs is that due to Lee and Mitchem [57] who showed that h(G) < (02 + 1) In1/21 . This was improved, independently, both by Zhikang Lu [60] and by McDiarmid and Luo Xinhua [67] who proved, respectively, that
h(G) < 20 [n 1/2j and
h(G) < 20(n -
1)1/2.
Both of these results are obtained by a sequential colouring method, in which the size of each colour class is kept below some limit t. It is then shown that there must always be a spare colour available to allow the next vertex to be coloured.
By a refinement of this method, limiting the degree sum of each colour class rather than its size, it was shown in [29] that
h(G) < 2(20m)1/2 + (20 - 1)L
(1)
which improves the previous bounds if G is not regular. In fact a slightly stronger result is proved. For any integer k, the k-core V (k) of a graph G is the set (possibly empty) of vertices which remain after vertices of degree less than k have been repeatly removed until the minimum degree is at least k. Then we have the following:
Theorem 3.1 Let G be a graph with m edges and maximum degree A. Then for any integer k > 2, h(G) < Max { IV(k) I , 2(2(k - 1)m)1/2 + (2k - 3)A}
.
For any graph, I V o+1) I = 0, which gives inequality (1) above. However for some classes of graphs, much better results can be obtained. For example, for any tree T, I V(2)1 = 0, hence
h(T) < 2(2m) 1/2 + A.
Similarly, for any planar graph I V(6)1 = 0, and for a graph of genus ry > 1, I V(7)1 < 12 (ry - 1). Thus we have:
Keith Edwards
18
Theorem 3.2 For any planar graph G with m edges, h(G) < (40m)1/2 + 90. For any graph G of genus ry > 1, with m edges h(G) < max {12(-y - 1), (48m) 1/2 + 110}
.
By a different technique, Krasikov and Roditty [55] improved the upper bound for general graphs to roughly A(2n)1/2. Their result does not however give improved bounds for the special classes above. 3.2
Lower bounds for h(G)
The only significant lower bound for h(G), apart from the easy ones given above, is due to Alon, and is given in [55]. He considered the probability of a fixed colouring of n vertices being a valid harmonious colouring for a random graph Gn P, for suitable p, and hence proved the following:
Theorem 3.3 Suppose that A satisfies 1041og2(2n) < A
(2 - s) log n/ log log n
provided n is large enough. He also noted that an example of Erdds shows that there is an irreducible graph Gn with n vertices for which
0(Gn) < logn/log2+2. Remark There are arbitrarily large graphs with achromatic number 2, for example the complete bipartite graphs Kn,n. However, as pointed out in [31],
the fact that v(k) is finite means that for any d, k, there are (ignoring isolated vertices) only finitely many graphs with maximum degree at most d and achromatic number at most k.
Keith Edwards
20
3.5
The values of h(G)
The harmonious chromatic number must take a value between the lower bound Q(m) and the number of vertices n. By considering trees formed from a star attached to the end of a path, Mitchem [73] showed that it can take any such value:
Theorem 3.4 Let n = m + 1 and let t satisfy Q(m) < t < n. Then there is a tree T with n vertices such that h(T) = t. For general graphs Mitchem proved that we can specify the number of vertices and the number of edges:
Theorem 3.5 Suppose that m satisfies 0 < m < (2), and t satisfies Q(m) < t < n, then there is a graph G with n vertices and m edges such that h(G) = t.
4
Special graphs
As with most graph parameters, there are a number of results giving more or less exact determination of the harmonious chromatic number for various special graphs. These rely mainly on ingenious ad hoc arguments based on the special structure of the graphs. 4.1
Paths and cycles
Naturally among the first graphs to be considered, for all three of the parameters h, A and i, were Pn and C, , respectively the path and cycle on n vertices. The value of the parameters for these graphs is equivalent to the existence of Eulerian graphs with a certain number of vertices and edges. For example, for h(C,,), it is easy to see that Cn can be harmoniously coloured with t colours if and only if there is an Eulerian (simple) graph on t vertices with n edges. By this argument, Mitchem [73] showed the following:
Theorem 4.1 If Q(n) is odd and R(n) 1, 2 then h(Cn) = Q(n). If Q(n) is even and R(n) > Q(n)/2 then h(Cn) = Q(n). Otherwise h(Cn) = Q(n) + 1. The corresponding result for paths, proved by Hopcroft and Krishnamoorthy [48] can be stated as follows:
Theorem 4.2 Let m = n - 1 be the number of edges of P. If Q(m) is odd, or if Q(m) is even and R(m) > Q(m)/2 - 1, then h(Pn) = Q(m). Otherwise h(Pn) = Q(m) + 1. Note that the harmonious chromatic number can fail to meet the lower bound Q(m) even for paths.
Harmonious Chromatic Number
21
Similar results for the line-distinguishing chromatic number of paths and cycles are given by Frank, Harary and Plantholt [33] and Al-Wahabi et al. [2], and for the achromatic number by Geller and Kronk [36], Bories [11], Bories and Jolivet [13] and Hare et al. [43]. Mitchem also showed that if G is the disjoint union of two cycles C, and
Cs, then h(G) = h(Cr+s). In the same spirit, Georges [38] determined completely the case of a collection of disjoint paths, generalising Hopcroft and Krishnamoorthy's result:
Theorem 4.3 Let G consist of r disjoint non-trivial paths with m edges in total. If Q(m) is odd, then h(G) = Q(m). IfQ(m) is even, then h(G) = Q(m) if r + R(m) > Q(m)/2, otherwise h(G) = Q(m) + 1. This result is also implicit in [14]. By considering the decomposition of various complete and complete bipartite graphs into cycles, Georges also gives a complete solution for collections of 4-cycles:
Theorem 4.4 Let bC4 be the graph consisting of b disjoint 4-cycles, and let m = 4b, the number of edges of the graph.
If Q(m) is even and R(m) > Q(m)/2, or Q(m) Q(m) Q(m) Q(m)
1(mod 8), or 3(mod 8) and R(m) > 3, or 5(mod 8) and R(m) > 6, or 7(mod 8) and R(m) > 5,
then h(bC4) = Q(m). Otherwise h(bC4) = Q(m) + 1. Furthermore, Georges established that h(bC4,.) = h(brC4). 4.2
Complete graphs
For a single complete graph K,,,, we of course have h(K,,) = n. If a graph G consists of a collection of disjoint complete graphs K('), ..., K(r), where K(') has ki vertices, then it is clear that G can be coloured with colours 1, 2, ... , C if and only if the set {1, 2, ... , C} has r subsets S1, ..., S, with JSiJ = ki, satisfying ISi n S,1 < 1, i # j. Although there are various partial results on when this is possible, it does not appear to be completely solved. However Georges proved the following result for the case where the number of complete graphs is relatively small.
Theorem 4.5 Let G be a graph consisting of disjoint complete graphs K(l), ... , K(r), where KW has ki vertices. If each ki > r - 1, then h(G) _
ki - (;).
Keith Edwards
22
If G is a complete bipartite graph K,.,,, then it is immediate that h(G) = r+s since G has diameter 2. For a graph consisting of 2 complete bipartite graphs Kp,q and Kr,,, where p = max {p, q, r, s}, Georges proved that
p+1
h(Kp,9UK,.,s) 4.3
if p> r+s
q+r+s if r+s>p.
Square grids
The square grid P x P,,, has m = 2n(n - 1) so that Q(m) = 2n. By an ingenious construction, Miller and Pritikin [72] show that h(Pn x Pn) = Q(m) when n is even, but for n odd they are only able to show that h(P,, x Pn) < Q(m) + 1. 4.4
Binary and r-ary trees
A number of authors have considered the problem of determining h for the complete r-ary tree of height H, which we denote T,,H. Miller and Pritikin [72]
considered the binary trees T2,H and showed that h(T2,H) = 0(2H/2), and Mitchem [73] gave some improved bounds for T2,H. Zhikang Lu [61] produced a more efficient colouring scheme to show that, for k > 3, h(T2,2k) < Q(m) + 2, and proved a similar result for ternary trees, showing that, for k > 2, h(T3,2k+l) < Q(m) + 4. In [63] he also gave some estimates for h(T4,H). Mitchem [73] solved the problem for trees of height 2, showing that h(Tr,2) = 13(r + 1)/21. It follows from the results in [26] and [19] that for each r > 2, h(T,.,H)
Q(m) and b(Tr,H) = q(m) if H is sufficiently large. Finally the author has recently verified that with one exception, all complete trees of height at least 3 have harmonious chromatic number Q(m). Thus we can state the following.
Theorem 4.6 Let Tr,H be the complete r-ary tree of height H. Then
h(Tr,1) = r + 1 h(T,.,2) _ [3(r + 1)/21 h(T2,3) = 7 (= Q(m) + 1)
h(T,.,H) = Q(m) otherwise. 4.5
Stars
For a single star with m edges Kl,,n, clearly h(K1,,n) = m + 1. The case of a graph which is a forest consisting of a number of stars is solved in [30]:
Theorem 4.7 Let F be a forest consisting oft stars of sizes m1 > Let m = i=1 mi, the number of edges of F. Then
> mt.
Et
Max
h(F) = max
> (mi + i) i-1
1: 1 < k < Q(m)
,
Q(m)
.
Harmonious Chromatic Number 4.6
23
Random graphs
It was noted in Section 2.1 that for any graph of diameter two, the harmonious chromatic number equals the number of vertices. Since almost all graphs have diameter two, it follows immediately that h(G) = IV(G) l for almost all graphs G. The achromatic number of a random graph is much less straightforward, but was determined quite precisely by McDiarmid [66], who showed the following:
Theorem 4.8 The proportion of graphs G with n vertices which satisfy
n/(k + 1) < z/'(G) < n/(k - 1), where k = (loge n)1!2, tends to 1 as n tends to infinity.
The upper bound derives from a fairly straightforward probability argument, while the lower bound is obtained from analysis of the behaviour of a very inefficient colouring algorithm, the "bounded sequential algorithm". 5
Asymptotic results
In their original paper on line-distinguishing chromatic number [33], Frank, Harary and Plantholt derive expressions for A (P,,) and A(C,,) and remark that
each of these is asymptotic to
2n. Since also
2n - Q(m) for a path or
cycle, we have
A(P.) - A(Cn) - Q(m). It is also easy to see from Theorems 4.1 and 4.2 that
h(Pn) - h(Cn) - Q(m). Although in this case the values of h(Pn) and h(Cn) are known precisely, there are many classes of graphs for which the exact determination of harmonious chromatic number appears to be very difficult. In most cases the only available lower bound is Q(m), so determining h(G) exactly is likely to be hard unless
h(G) = Q(m), and even if equality holds, proving so requires giving (or at least proving the existence of) a colouring which uses all but a very few of the available colour pairs. Describing such a colouring is likely to be very difficult unless G has a lot of structure. We might however expect to have more success in showing that for certain classes of graphs, h(G) N Q(m). Of course, being
a weaker statement, it is more likely to be true, but the fact that it is less precise also means that techniques (e.g. probabilistic) which "waste" a small proportion of colour pairs may still succeed.
Keith Edwards
24
Notation If F is a class of graphs, we will use the shorthand
h(G) - Q(m) for G E F to mean: For any E > 0, there is an integer M such that if G is a graph in the class r, and G has m > M edges, then Q(m) < h(G) < (1 + e)Q(m). Other asymptotic results follow from the theorem of Wilson [82] on the decomposition of complete graphs into copies of a graph:
Theorem 5.1 Given a graph G, there exist subgraphs G1, ..., Gt of the complete graph & such that each edge of K,, belongs to exactly one of the graphs Gi and such that each Gi is isomorphic to G provided that (i) n is sufficiently large, (ii) the number of edges of G divides (2), and (iii) the greatest common divisor of the degrees of the vertices of G divides n - 1.
This theorem can be interpreted as giving sufficient conditions for the graph tG consisting of t disjoint copies of some graph G to have harmonious chromatic number n. As described in [29], it follows from this that for any fixed graph H, if we consider the class of graphs G consisting of disjoint copies of H, i.e. G = kH for some k, we have h(G) - Q(m). More generally if k is a positive integer and F(k) is the class of all graphs whose components have at most k vertices, then h(G) - Q(m) for graphs G in F(k). In [29] this is extended to a considerably wider class of graphs, as we see in the next section. 5.1
Fragmentable graphs
Definition A class r of graphs is said to be fragmentable if. For any E > 0,
there are positive integers no, c, depending on e, such that if G E r is a graph with n > no non-isolated vertices, then there is a set S of vertices, with ISI < en such that each component of G - S has at most c vertices. Fragmentable classes of graphs include trees, planar graphs and, more generally, graphs of genus at most -y for some fixed -y > 0. They also include some classes of unbounded genus such as d-dimensional grids for fixed d > 3.
Suppose r is a fragmentable class all of whose elements have maximum degree at most d. Let G be an element of r and suppose that G has a set of vertices S of size En which fragments G into components of size at most c(E). It follows from the result above that we can find a harmonious colouring of all of these components using at most (1 + E)Q(m) colours. We then use the following lemma, proved in [28]:
Lemma 5.2 Let d be a fixed integer, and let e > 0. Suppose that G is a graph with maximum degree at most d, n vertices and m edges, and G has a partial harmonious colouring with k colours such that at least (1 - e)n of the vertices are coloured. Then h(G) < k + 12d2e"2n1/2 provided n is large enough.
Harmonious Chromatic Number
25
This lemma shows that, as we would expect, we can extend the colouring to the small remaining set of vertices S with only a small number of extra colours (although it may be necessary to recolour some already coloured vertices).
Therefore h(G) - Q(m) for graphs G in F. In particular this shows that bounded degree planar graphs can be coloured near-optimally. 5.2
Bounded degree graphs
Finally in [28] is it shown that for any fixed d, h(G) - Q(m) for G in the class Fd of graphs of maximum degree at most d. We give a rough outline of the proof below. First consider the case when G is a d-regular bipartite graph with the vertex set V having parts A and B, and suppose we wish to harmoniously colour G using disjoint colour sets for A and B. If G has m edges, and we use a colours for the set A, and b colours for B, then we must have ab > m. Thus the total number of colours, a+ b, must be at least 2V-m. We will colour G so that each
of a and b is about \. and for each vertex of A, indeChoose a colour set CA of size about pendently choose a colour uniformly from CA, and use it to colour the vertex. A minor adjustment of the resulting colouring of A may be needed to ensure that no two vertices with a common neighbour receive the same colour; we omit the details of this. With high probability the degree sum of each colour since the total degree of the vertices will be roughly the same, i.e. about in A is m.
We now form a d-uniform hypergraph H. The vertex set V(H) is equal to CA, the set of colours used on A. The edges of H correspond to vertices of G
as follows: for any vertex v E B, the hypergraph H has an edge e containing the d colours used on the neighbours of v. The degree dH(x) of a vertex x in the hypergraph H is determined as follows: x is an element of the colour set CA, and the degree dH(x) is the degree sum of the colour x in G, which is about N/m-, as noted above. Hence H is approximately regular of degree about \,Fm-. For
any two vertices x, y of H, the number of edges in H containing x and y, called the codegree of x and y, is likely to be small in comparison with Thus we have exactly the situation in which to apply the Pippenger-Spencer theorem [75] on the chromatic index of uniform hypergraphs. This gives an -,/-M.
edge colouring of the hypergraph with about \ colours. Transferring the colours from the edges e to the corresponding vertices v of B completes a colouring c of G. This colouring has disjoint colour sets, of size about \/M_, for each of A and B. Furthermore it is a harmonious colouring, for suppose that (v, w), (v', w') are two distinct edges in G, where v, v' E A, w, w' E B. If c(w) # c(w'), then the two edges have distinct colour pairs. So suppose
c(w) = c(w'). Then if w = w', the colours c(v) and c(v') must be distinct, since we ensured that any two vertices in A with a common neighbour received different colours. So suppose to # w'. Then since c(w) = c(w'), we know that
Keith Edwards
26
e,,,
,
e,,,' received the same colour in the colouring of H, hence e,,, and e,,,, must
be disjoint. But c(v) E e,, and c(v') E e,,,', so we have c(v) # c(v'). Hence distinct edges have distinct colour pairs, and the colouring is harmonious.
Now suppose that G, rather than being a bipartite graph, is a d-regular r-partite graph, with the r parts all having the same number of vertices. In this case we can do much the same as above, in r stages. At each stage one of the r parts of the graph is coloured. The first part is coloured randomly, each subsequent stage uses the Pippenger-Spencer theorem to extend the colouring to another part. The total degree of each part is 2m/r, and we will use about (2m) 1/2/(r - 1) colours for each part. Each colour thus has degree sum about (2m) 1/2(r - 1)/r. In order to construct the hypergraph at each stage we have to assign, randomly, some temporary colours to the uncoloured vertices; these colours are discarded once the stage is complete. The hypergraph has common degree about (2m)1/2/r, so the Pippenger-Spencer theorem allows us to colour
it with about (2m) 1/2/(r - 1) colours as required, allowing the process to continue. In total we use about (2m)1/2r/(r - 1) colours. Provided r is large, this is close to Q(m). Finally suppose we have any d-regular graph G. We choose an integer r,
and randomly assign the vertices of G to one of r "parts". Of course, we do not truly obtain an r-partite graph, but it is not hard to see that with high probability nearly all the edges of G will join vertices in distinct parts. Thus the graph can be regarded as nearly r-partite. We therefore colour it as above. This may not give a true harmonious colouring, because of edges whose endpoints are in the same part. However there are few such endpoints, and these can be recoloured using the extension Lemma 5.2. Provided we choose r large enough, this suffices to give a colouring of G with roughly Q(m) colours. The extension to all graphs of maximum degree at most d is straightforward. For details of the proof see [28]. In [19], the equivalent result for achromatic number is given, namely that V(G) - q(m) for G in r'd. (Of course q(m) and Q(m) are asymptotically equivalent.)
6
Trees
For most kinds of graph colouring, questions relating to the colouring of trees are usually answered easily, and in many cases are trivial. However, this is far from being the case for harmonious and complete colourings, for, as we shall see in Section 7, these colourings give rise to NP-hard problems on trees. Nonetheless, the simple structure of trees does allow us to obtain better bounds than for most other classes of graph, and in some cases to obtain exact results.
Harmonious Chromatic Number 6.1
27
General bounds
For achromatic number, the best lower bound for trees is that of Farber, Hahn, Hell and Miller [31]. They show that extending a tree by a fairly small number of edges must increase the achromatic number, and so obtain a recurrence which leads to an upper bound for the number of edges of a tree with maximum degree at most d and achromatic number at most k: Theorem 6.1 Let T be a tree with m edges, maximum degree at most d, and satisfying O(T) < k. Then
m
d.
They then prove the following corollary:
Theorem 6.2 Let T be a tree with m edges and maximum degree A. Then L2 + (m -
\A2 1)
-
4)1/2J :5 N(T)
q(m).
Note that for fixed A the lower bound here is asymptotic to m1/2
q(m)//.
For harmonious colouring, McDiarmid and Luo Xinhua [67] noted that for
a tree T, h(T) _< 2(On)'/2. They also noted that for complete r-ary trees it could be shown that h(T) < 2(2n)1/2, and asked if it could be proved that h(T) = O(nl/2 + A) for all trees. As mentioned in Section 3.1, it was proved in [29] that for any tree T with m edges, h(T) < 2(2m)1/2 + A < 2Q(m) + A.
Note that this is asymptotic to 2Q(m) and so appears to be inferior to the bound for V)(T) given above. 6.2
Exact results
If we wish to colour a graph G harmoniously with C colours 1, . . . , C, we have to colour the vertices of G such that each colour pair {i, j} is used at most once. It is therefore natural to think of these colour pairs as being (represented by) the edges of the complete graph on the vertices 1, . . . , C. Now consider the situation which arises when we have a partial colouring of a graph G (for example at some stage in a sequential colouring process). Assuming that we
are not going to recolour any vertices already coloured, then any coloured vertex whose neighbours are also coloured is, in a sense, "inactive". Thus it is helpful to consider the "active" part of G, namely the subgraph A induced by the uncoloured vertices and their coloured neighbours, together with the
28
Keith Edwards
partial colouring of A. It is also helpful to think of the colour pairs which do not yet appear on any edge of G forming the edges of a graph on the vertices 1, . . . , C. We call this graph the unused colour pairs graph U. As the colouring of the graph progresses, the graphs A and U evolve. Now consider the special case of an exact graph G where m is exactly (2) .
(In fact this case is not so special since we can always add (2) - m extra isolated edges to the graph, without affecting the value of h(G).) In this case we wish both A and U to evolve away to nothing. Unless G is very highly structured this is likely to be difficult to achieve, as we have to match up the colour pairs and the edges precisely. One approach is to try to colour a successively larger portion of G in such a way as to achieve progressively more structure in A and U. Eventually they may become sufficiently well structured that a single technique will complete the colouring all at once. Two such techniques were used in [26, 27] to prove some exact results on the harmonious chromatic number of bounded degree trees. First consider a tree T of maximum degree at most d having at least en leaves, where e is some small positive constant. Also suppose that T has exactly (2) edges, and we wish to colour it with C colours. As observed in Section 5.1, trees are fragmentable, so by removing some small set S of the vertices, where ISO = rin and ri is much smaller than e, we break up the tree into components all of size less than some constant c(ri). We then remove the leaves from all these components. Then as in Section 5.1, we colour these components in a highly symmetrical way. Thus at this stage the graph U of unused colour pairs is also highly symmetrical. We then extend the colouring to the set S. This to some extent destroys the symmetry of U, but not too much because ISI is much smaller than the number of edges in U. Note that because the tree without its leaves has only (1 - e) (2) edges, then by the asymptotic result of Section 5.1 we are able to do this colouring with only C colours, provided the tree is large enough. Now the active part of the tree, A, consists of a number of stars, namely the deleted leaves and their neighbours. The centres of the stars are coloured, and
the leaves are uncoloured. This is exactly the kind of simple structure for A that is desirable. Let the total degree of the star centres which are coloured i be ri. It is not hard to see that completing the harmonious colouring by colouring A with the unused colour pairs of U is exactly equivalent to orienting the edges of U so that the outdegree of vertex i is exactly ri for each colour i. Using Hall's theorem we can derive sufficient conditions for this, and in [26] it is shown that it is always possible provided the tree is large enough. Thus we have the following:
Theorem 6.3 Let d be a positive integer and let e > 0. Then there exists an integer M = M(d, e) such that if T is any tree with m > M edges, maximum
Harmonious Chromatic Number
29
degree at most d and at least en leaves, where n = m + 1, then T satisfies h(T) = Q(m). Now consider trees which have very few leaves. Since the average degree of a tree is less than 2, almost all the vertices have degree 2, so the tree consists largely of long paths. Note that there is no hope of proving the above result
for these trees, for as noted in Section 4.1, arbitrarily long paths P,' have h(P,,,) = Q(m) + 1. More generally, if T has (2) edges where C is even, then it will be impossible to colour T with C colours unless T has at least C vertices of odd degree. This is because the degree sum of each colour must be C - 1 which is odd, thus there must be at least one odd degree vertex of each colour. In what follows we will largely ignore such parity problems, and state simply that in some cases one extra colour will be needed. To colour T we first colour the vertices of odd degree (plus possibly some even degree vertices). To do this we use Theorem 6.3 above, extended to forests. After this, the active part A consists of a number of paths, with their endpoints coloured, but the rest of the vertices uncoloured. The paths may be of many different lengths. We now partition the colour set into 4 sets W, X, Y, Z, where IXI = IYI and IXI, IYI, IZI are all even. By colouring parts of the paths, we gradually improve the structure of both A and U, until we have used up all colour pairs except those in X x Z and Y x Z, so that U is a complete bipartite graph, and the active graph A consists entirely of paths of length 4. Furthermore, there are exactly 1 IXI IZI of these paths, with IXI IZI endpoints, and each colour in X occurs on exactly IZI of these endpoints (but the colours are arranged in an arbitrary way). We then colour the midpoints of all the paths with a colour from Y, so that each colour from Y occurs the same number (i.e. 1 IZI) of times. Now A consists of IXI IZI paths of length 2, each with an X-colour at one end and a Y-colour at the other, and each colour occurring exactly IZI times. From this we construct a bipartite multigraph M with vertex set X UY, and an edge joining x E X and y E Y for each path in A which has endpoints coloured x, y. Now M is regular of degree IZI, and hence can be edge coloured with IZI colours, i.e. those in the set Z. Transferring the colour of an edge to the midpoint of the corresponding path in A completes the colouring, using for A exactly the colour pairs in X x Z and Y x Z. Combining this result with Theorem 6.3 gives:
Theorem 6.4 Let d be a fixed positive integer. Then there is an integer N(d) such that if F is any forest with maximum degree at most d and with m > N(d) edges, then h(F) = Q(m) or Q(m) + 1.
Furthermore, if F has a least Q(m) - 2R(m) vertices of odd degree, then h(F) = Q(m).
Keith Edwards
30
The achromatic number counterpart of this is the following, proved in [19]:
Theorem 6.5 Let d be a fixed positive integer. Then there is an integer N(d) such that if F is any forest with maximum degree at most d, and with m > N(d) edges, then 4'(F) = q(m) or q(m) - 1.
Furthermore, if F has at least q(m) - 2r(m) vertices of odd degree, then b(F) = q(m). 6.3
Almost all trees
There are many results of the form: almost all graphs G have property P, meaning that the proportion of (unlabelled) graphs on n vertices which have property P tends to 1 as n tends to infinity. In most cases it is known that there are infinitely many graphs which do not have property P. In contrast there appear to be very few results of the form: almost all (unlabelled) trees T
have property P. The only general technique for showing such a result is apparently the "forbidden limb" technique of Schwenk [79]. A limb of a tree
at a vertex v is formed as follows: Delete some, but not all of the edges incident to v; the component containing v is a limb at v. Schwenk proved that for any given limb R, almost all trees T contain R as a limb (at some vertex). In fact he showed that if rn is the number of trees on n vertices not containing R as a limb, and to is the total number of trees on n vertices, then the growth rate of rn is strictly less than that of tn. Corollaries of Schwenk's theorem include the well known result that almost all trees have a non-trivial automorphism, and that almost all trees are cospectral, i.e. share the same characteristic polynomial with some other tree. There are also a number of extensions of the latter result. These are all proved by showing that if a tree contains some particular limb R, then it has the desired property, and then appealing to Schwenk's theorem. In [26] the following result is proved:
Theorem 6.6 Almost all trees T satisfy h(T) = Q(m). This result was conjectured (in a slightly different form) by Frank, Harary and Plantholt in the first harmonious colouring paper [33]. The method used to prove this result is quite different from the forbidden limb technique above. We first prove a slightly stronger version of Theorem 6.3, which shows that a sufficiently large tree which is "nearly" bounded degree satisfies h(T) = Q(m). Nearly bounded degree means that the maximum degree is not too big (O(logn)), and the total degree of the vertices of degree greater than some d is very small (< n/2 d). We then show that almost all trees are nearly bounded degree. Together with the fact that almost all trees have plenty of leaves, this gives the result.
Harmonious Chromatic Number
31
The corresponding result that '(T) = q(m) for almost all trees is proved in [19].
Remark It should be noted that the set of trees for which h(T) Q(m) is quite large. If sn is the number of such trees with n vertices, and to is, as above, the number of trees on n vertices, then although sn/t,,, -4 0 as n -+ oo, the growth rate of sn is equal to that of tn. This is easily seen by considering the number of trees with maximum degree greater than Q(m). If follows that Theorem 6.6 cannot be proved by the forbidden limb technique. 6.4
Algorithms
It follows from Theorem 6.4 that there is a simple, though not very practical, algorithm which will determine the harmonious chromatic number of a bounded degree tree to within 1 in linear time. The algorithm simply determines if the number of edges of the tree is at least N(d), if so it declares that h(T) is Q(m) or Q(m) + 1, otherwise it uses exhaustive search to determine Q(m). In fact, with a little more care, it is possible to determine h(T) exactly in polynomial time. Thus we have
Theorem 6.7 Let d be a positive integer. Then there is a polynomial time algorithm which will determine the harmonious chromatic number h(F) of a forest F with maximum degree at most d. In fact we first deal with the case of an exact forest, i.e. one with exactly (Qin`l) edges. The general problem for harmonious chromatic number reduces z m isoto this, since any forest can be made exact by adding R(m) _ (Q lated edges to F, without changing the value of h(F). The achromatic number problem is somewhat more tricky, since in general removing r(m) edges to make the forest exact will reduce the value of 0. Nevertheless, the corresponding result for achromatic number is given in [19], although the algorithm is considerably more complicated.
Theorem 6.8 Let d be a positive integer. Then there is a polynomial time algorithm which will determine the achromatic number J'(F) of a forest F with maximum degree at most d.
7
Computational complexity
As might be expected, the determination of h(G) and of z/)(G) are both NPhard problems. Yannakakis and Gavril [84] showed that the ACHROMATIC NUMBER problem
Instance Graph G, integer k. Question Is VY(G) > k?
Keith Edwards
32
is NP-complete even for the complements of bipartite graphs. This is one of a number of results which follow from their proof that the edge dominating set problem is NP-complete for bipartite graphs of maximum degree 3. Subsequently, ACHROMATIC NUMBER was shown to be NP-complete for bipartite graphs by Farber, Hahn, Hell and Miller [31], using a reduction from PARTITION. The line-distinguishing chromatic number .A(G) was shown to be NP-hard by Hopcroft and Krishnamoorthy [48], and a very short proof by D. S. Johnson
of the same result is also given in the paper. This proof is easily adapted for harmonious chromatic number. Consider the problem HARMONIOUS COLOURING:
Instance Graph G, integer k. Question Is h(G) < k? The following proof, based on Johnson's, shows that this NP-complete: HARMONIOUS COLOURING is obviously in NP. For completeness, reduce from INDEPENDENT SET. Let G, k be an instance of INDEPENDENT SET. Construct an instance G', k' of HARMONIOUS COLOURING as follows: G' consists of a component equal to G with one extra vertex u joined to all the other vertices, and a second component which is a k-clique. Set k' = IV (G) I + 1. Note that any colouring of the first component must use k' colours, one for each vertex. Then it is easy to see that a harmonious colouring of G' with k' colours exists if and only if G has an independent set of size k.
However as early as the paper by Frank, Harary and Plantholt [33], it was conjectured that the problem is hard for more limited classes of graphs. In this paper, a directed version of the problem is defined; thus in a harmonious colouring of a directed graph we are allowed, for any pair of colours i, j, both an i -> j edge and a j -4 i edge. It is pointed out in [33] that the problem of deciding if a directed graph can be harmoniously coloured with k colours is NP-complete even for graphs in which each component is a star with all edges directed outwards. This is done by a reduction from BIN PACKING, roughly as follows: An instance of BIN PACKING consists of a set of k bins, each of capacity B and a collection of objects ui, i = 1,. .. , n, each with an integer size s(ui). A solution to the instance is an arrangement of the objects in the bins so that each is exactly full. BIN PACKING is strongly NP-complete (see for example [34]), thus the size of the integers can be assumed to be in unary. It is also easy to restrict to instances for which k = B + 1. Then we reduce to DIRECTED HARMONIOUS COLOURING by constructing a graph consisting of n outwardly directed stars, of sizes s(ul), . . . , s(un). The colour of the centre of a star corresponds to the choice of bin in which to place the corresponding object.
Harmonious Chromatic Number
33
Frank, Harary and Plantholt conjectured that the undirected version of HARMONIOUS COLOURING is NP-complete when restricted to forests. This is proved in [30] by a more sophisticated bin-packing argument; in fact the paper shows that it is NP-complete to determine if a tree with (2) edges has harmonious chromatic number k. It is also shown that the undirected version can be solved in polynomial time if each component is a star. A rather similar reduction, this time from 3-PARTITION, is used by Bodlaender [8] to show that ACHROMATIC NUMBER is NP-complete even for graphs which are simultaneously cographs and interval graphs. Since the graphs constructed have (2) edges, it follows immediately that the HARMONIOUS COLOURING problem is NP-complete for these graphs also. Likewise it follows from the NP-completeness of HARMONIOUS COLOURING for trees, mentioned above [30], that ACHROMATIC NUMBER is also NP-complete for trees [19], answering a question originally posed by Hedetniemi, Hedetniemi and Beyer [44] and repeated by a number of other authors [8, 21, 31, 52].
It is not however the case that the two problems are equivalent in complexity for all classes of graphs. We can give examples of classes of graphs for which one problem is in P, while the other is NP-complete.
Example 7.1 Let G be any graph, and let G' be the graph formed from G by adding a single new vertex u and joining u to each vertex of G. Then it is easy to see that b(G') = 1(G) + 1, so the problem of determining O(G') is the same as that of determining &(G). However all such graphs G' have diameter at most 2. Since ACHROMATIC NUMBER is NP-complete, and graphs of diameter 2 are easily recognised, it follows that ACHROMATIC NUMBER is still NP-complete when restricted to graphs of diameter at most 2.
On the other hand, for any such graph G, h(G) = IV(G)I, so the problem HARMONIOUS COLOURING is trivial, and belongs to P.
Example 7.2 Consider the class of graphs known as split graphs, whose ver-
tex set V can be partitioned into two sets A and B, such that A induces a complete graph and B a null graph. It is easy to recognise split graphs and find the sets A and B by looking at the degrees of the vertices. Now the achromatic number of any split graph is Al I+ 1 if the vertices of B are collectively adjacent to every vertex of A, and IAA otherwise. Clearly this can be determined in polynomial time. Thus the ACHROMATIC NUMBER problem for split graphs belongs to P. However the HARMONIOUS COLOURING problem for split graphs is NP-complete. This can be seen by the following reduction from CHROMATIC NUMBER. Let G = (V, E) be a graph. We construct a split graph G'. The vertex set of G' is E U V. The edge set of G' is as follows: for each (u, v) E E, there is an edge in G' joining the vertex
(u, v) E E to the vertices u and v in V. In addition, E induces a complete graph in G', and V induces a null graph in G'. Then it is easy to see that h(G') < JEl + k if and only if X(G) < k, and the result follows.
Keith Edwards
34
This example also shows that HARMONIOUS COLOURING is NP-complete for graphs of diameter 3, since G' will have diameter 3 provided G has no isolated vertices. (I am indebted to Niall Cairnie for this example.) Fixed number of colours It should be noted that both HARMONIOUS COLOURING and ACHROMATIC NUMBER can be solved in polynomial time if the number of colours k is fixed. For HARMONIOUS COLOURING this is obvious, since if the graph
has more than (2) edges then it cannot be coloured with k or fewer colours, so the answer is "no". If the graph has (2) or fewer edges then we can do an exhaustive search. For ACHROMATIC NUMBER the result is slightly less straightforward, but was proved by Yannakakis and Gavril [84] as follows.
Suppose that a graph G satisfies '(G) > k. Then there is a set of (2) edges in G which uses all the colour pairs. These edges have between them at most k(k - 1) endpoints. Thus there is an induced subgraph H of G with at most k(k-1) vertices such that '(H) > k. On the other hand, if such a subgraph H exists, then '(G) > k also. So to determine if O(G) > k we simply form each induced subgraph of G on at most k(k - 1) vertices and check each to see if its achromatic number is at least k. Since the number of such induced subgraphs is at most (k(k 1)) and k is fixed, the algorithm is polynomial. A more efficient 0(m) algorithm is given in [31] and relies on the results of Hell and Miller given in Section 3.4. Even this algorithm, however, is only practicable for very small k, because it uses the value of the function v(k) defined in Section 3.4. Of course given that the ACHROMATIC NUMBER problem is NP-complete when k is unbounded, it is likely that any algorithm will have time complexity exponential in k. Approximation algorithms An approximation algorithm is one which delivers an approximate solu-
tion to a problem. For a maximisation problem such as the determination of achromatic number, an algorithm has approximation ratio a if it always produces a solution whose value is at least of the optimum. Chaudhary and Vishwanathan [21] give a polynomial time approximation algorithm for the achromatic number with approximation ratio 0(n/ logn). For graphs of girth at least 7, they give a simple algorithm with approximation ratio 0(n7/20).
8 8.1
Related topics Parameters related to achromatic number
Several parameters similar to the achromatic number have been studied. The most closely related is the pseudoachromatic number. A pseudocomplete colouring is a colouring of the vertices, such that for any pair of colours, there
Harmonious Chromatic Number
35
is at least one edge whose endpoints are coloured with this pair of colours. It differs from a complete colouring in that it need not be proper, i.e. the graph induced by a colour class need not be a null graph. The pseudoachromatic number V)s (G), is the greatest number of colours in a pseudocomplete colouring
of G. The pseudoachromatic number was introduced by Gupta [39] and has also been studied by Bhave [7], Sampathkumar and Bhave [78] and Bollobas, Reed and Thomason [10]. Another related concept is the ajointed number defined by Cook and Evans [25]. A simple fold is a homomorphism which identifies a pair of nonadjacent vertices having a common neighbour, and a fold is a sequence of simple folds. The ajointed number aj(G) of a connected graph G is the largest n for which there is a fold onto K. It is easy to see that x(G) < aj(G) < O(G) < Os(G) for any connected graph G. Further results on aj(G) are given in [18]. Bollobas, Catlin and Erd6s [9] defined the contraction clique number ccl(G) to be the maximum number of colours in a pseudocomplete colouring of G, such that the graph induced by each colour set is connected. Clearly ccl(G) < 0, (G)
for any graph G. Hadwiger's conjecture is that ccl(G) > x(G) for any G. For this reason ccl(G) has also been called the Hadwiger number of G. 8.2
r-reduction number In Section 1 we described how the harmonious chromatic number of a graph
could be defined as the number of vertices in the smallest graph obtainable from G by repeatedly identifying vertices at distance at least 3. This suggests a natural generalisation.
Definition Let G be a (not necessarily simple) graph. Suppose that u, v are two vertices of G satisfying dG(u, v) > r. Form a new graph G' by identifying
u and v to form a new vertex uv, so that each edge incident with u or v becomes an edge incident with uv. We will call this operation an elementary r-reduction.
Definition An r-reduction is a sequence of (zero or more) elementary rreductions. We will also say that H is an r-reduction of G if H is obtained from G by an r-reduction. Definition Let G be a (not necessarily simple) graph. The r-reduction number X(r) (G) is the least integer k such that G has an r-reduction with k vertices. It follows from the discussion above that XP) (G) is equal to the harmonious chromatic number h(G), and it is also easy to see that x(2)(G) is the ordinary chromatic number x(G). What can be said about x(r), r > 4? We first make some simple observations:
1. If G is a simple graph then so is any r-reduction of G, r > 3, so we can restrict our attention to simple graphs.
Keith Edwards
36
2. For any graph G, x(4) (G) > h(G) > x(G), and in general we have for each r > 2, X1(G) > x(T )(G)
3. If G is connected, and r > diam(G), then X(') (G) = IV(G)I. More generally, if G has k components, then for all sufficiently large r, x(r) (G) = I V (G) I - k + 1.
4. If G has girth at least r, then so does any r-reduction of G. 8.3
The 4-reduction number x(4)
A special case of observation 4 above is that for any bipartite graph G, any 4-reduction of G also has girth at least 4, i.e. is triangle free. Thus if G has more than k2/4 edges, any 4-reduction of G has at least k vertices, i.e. x(4) (G) > k. Thus we have that for bipartite graphs G, x(4) (G) > 2vfm-.
First consider paths and cycles. If m = t2 for some even integer t, then it is easy to see that x(4)(Pm+i) = x(4)(Cm) = 2\. For the complete bipartite graph Kt,t is Eulerian, and hence has a closed Eulerian trail of length m. Thus we can label the vertices of Pm+1 or Cm with the names of the vertices of Kt,t
in the order in which they appear in the Eulerian trail, and then perform a sequence of elementary 4-reductions joining any two vertices with the same label until we obtain Kt,t. In general it is clear that x(') (Cm) < k if and only if there is an Eulerian graph on k vertices with m edges and girth at least r. A similar statement holds for paths. However, little is known about how many edges a graph with a given number of vertices and given girth can have. We now consider bipartite graphs. In Section 5.2 we sketched a proof that a d-regular bipartite graph with m edges could be harmoniously coloured with roughly 2j colours, using distinct colour sets of size about for the two parts of the graph. This proof shows that for such graphs, x(4) is approximately 2V-m. The result is easily extended to bipartite graphs of maximum degree at most d, hence we have that for bipartite graphs G of maximum degree at most d, X(4) (G) - 2vrm-. NP-hardness of x(4)
We consider the problem 4-REDUCTION NUMBER:
Instance Graph G, integer r. Question Is x(4) (G) < r?
Harmonious Chromatic Number
37
We will show that this problem is NP-complete even when restricted to graphs in which each component is a star. As for the DIRECTED HARMONIOUS COLOURING problem considered in Section 7, we reduce from the strongly NP-complete problem BIN PACKING. Recall that an instance of BIN PACK-
ING consists of a set of k bins, each of capacity B and a collection U of objects ui, i = 1, . . . , n, each with an integer size s(ui). We can assume that the sum of the sizes E 1 s(ui) is exactly kB, since otherwise a solution is clearly impossible. We can also assume that each s(ui) > k, for otherwise we can multiply each s(ui), and B, by k + 1 without altering the solution to the problem. T o form our graph G, we take a star of size s(ui) for each i = 1, ... , n,
and B - k stars of size B. Take r = 2B. Now since G is bipartite and has B2 edges, it follows from above that x(4) (G) > 2B, and that x(4) (G) = 2B if and only if G has a 4-reduction which is a triangle free graph on 2B vertices with B2 edges. It is well known that the only such graph is KB,B. Hence x(4)(G) = 2B if and only if the stars can be packed edge-disjointly into KB,B. It is clear that the centres of all the stars of size B must go on the same side of KB,B. Hence having packed these stars, we then have to pack the remainder into Kk,B. Since each s(ui) > k, the centres of these stars must also go on the same side of KB,B, i.e. each centre must be identified with one of the k vertices. This is clearly possible if and only if the instance of BIN PACKING has a solution. Note that HARMONIOUS COLOURING is solvable in polynomial time for this class of graphs. 8.4
Applications
Although ordinary vertex colourings have many applications, this does not appear to be the case for harmonious or complete colourings. However, Cichelli
[24] used a idea similar to harmonious colouring to implement perfect hash functions, and it appears that harmonious colourings could have applications to some forms of data compression. The idea is as follows: Suppose that we wish to store a sparse graph efficiently, while still allowing very fast verification of adjacencies. Suppose we can harmoniously colour the graph efficiently, that
is with C colours where C is about Q(m). We can store the graph using a C x C matrix (actually only half of it need be stored since it will be symmetric)
and a lookup table for the colours of the vertices. The (i, j)th entry of the matrix will contain the entry (u, v), the unique edge with endpoints coloured i, j (or a special marker if there is no such edge).
In order to determine if u and v are adjacent in the graph, we look up the colours of u and v, say i and j respectively, and then check whether the (i, j)th entry of the matrix is (u, v). Thus checking the adjacency requires only a very small number of simple steps and is constant time. Also the total storage space used is about n log n bits for the lookup table and 2m log n for the matrix. This is about the same as for an adjacency list representation,
Keith Edwards
38
which does not however have constant time adjacency checking. In practice this is most likely to be useful in cases where an input stream consists of pairs u, v, and it has to be checked that the pairs satisfy some fixed relation. 9 9.1
Open problems General bounds
The upper bounds for h(G) given in Section 3.1 are almost certainly not the best possible. It was suggested by Krasikov and Roditty [55] that the right upper bound for general graphs is around 0vfn-. As mentioned in Section 3.2, the only non-trivial lower bound is that of Alon [55]. It seems that a substantially new idea is needed to obtain further lower bounds. Apart from the results for irreducible graphs, there are few bounds of note for the achromatic number. 9.2
Special graphs
Open problems on special graphs include the complete determination of h for collections of cycles, and, more generally, for all graphs of maximum degree at most 2, and for all collections of complete graphs. For achromatic number,
there appear to be only a few results on special graphs apart from those for paths and cycles, though presumably analogues of all the results in Section 4 could be obtained. 9.3
Trees
Although more is known about the harmonious chromatic number of trees
than of other graphs, there are still some notable open problems. The best known upper bound for general trees is surely too big; it should be possible to
prove that h(T) is no more than about /Q(m). In fact, it seems quite likely that h(T) < Q(m) + A for all trees T. 9.4
Bounded degree graphs
As above, let rd be the class of graphs with maximum degree at most d. We know that h(G) - Q(m) for G in I'd, as described in Section 5.2, and that h(T) < Q(m) + 1 for all sufficiently large trees in Pd. Between these results is a large gap in our knowledge of the behaviour of h(G) for bounded degree graphs. Let Cd be the quantity defined by Cd = max {(h(G) - Q(m)) : G E Pd}
Perhaps the most important question is the following:
.
Harmonious Chromatic Number
39
Question 9.1 Is cd bounded for each positive integer d? The example of a d-star shows that Cd cannot be bounded independently of d. Further, if d = 4r + 2, r > 1, then there are infinitely many G E I'd
such that h(G) - Q(m) > d/2. For if k is any positive integer, there is a d-regular graph with (k + 2)((k + 2)d - 1) vertices. Then Q(m) = (k + however since each degree sum must be a multiple of d, it is not hard to see that h(G) > (k + 1)d. Hence cd > d/2. A similar result holds for other values
2)d,
of d.
We can also restrict attention to, say, planar graphs. Let pd be the quantity defined by
pd = max {(h(G) - Q(m)) : G E I'd, G planar}. Question 9.2 Is Pd bounded for each positive integer d? The example of a d-star above shows that pd also cannot be bounded independently of d, however it seems possible that there is an absolute constant A, such that for any d, h(G) < Q(m) + A for all sufficiently large planar G E Fd. Finally, let d be a fixed integer. Then if G is a d regular graph on k(kd+ 1) vertices, we have m = (kd2 1), so Q(m) = kd + 1. It seems possible that in this case h(G) = Q(m) provided only that G is sufficiently large, although this would be a very strong result. 9.5
Algorithms and complexity
A major open problem is the complexity of HARMONIOUS COLOURING and ACHROMATIC NUMBER for graphs of bounded degree. It seems likely that both problems can be solved in polynomial time for graphs of degree at most 2, but this has not been established. Conversely, for k > 3, the problems for graphs of degree at most k seem likely to be NP-complete, although for trees of degrees at most k, they can be solved in polynomial time, as described in Section 6.4. However all current NP-completeness proofs for these problems rely on the presence of vertices of large degree. Another open problem concerns the complexity of the following problem:
Instance Graph G with m edges.
Question Is h(G) = Q(m)? We will call this OPTIMAL HARMONIOUS COLOURING. Clearly this problem is just a special case of HARMONIOUS COLOURING in which we set
k = Q(m). It remains NP-complete for general graphs and for trees, but it is not known to be NP-complete for regular graphs for example (HARMONIOUS COLOURING is NP-complete for regular graphs).
Keith Edwards
40
Acknowledgements I am grateful to Niall Cairnie and the referee for a number of useful comments on this paper. Remarks on references
The following is intended to be a comprehensive list of references concerning the achromatic number, harmonious chromatic number and line-distinguishing chromatic number. It does not include papers concerned with edge colouring parameters, such as the achromatic index and the point-distinguishing chromatic number. Also omitted are papers concerned only with hypergraphs.
References [1]
J. Akiyama, F. Harary & P. Ostrand, A graph and its complement with specified properties VI: Chromatic and achromatic numbers, Pacific Journal of Mathematics, 104 (1983), 15-27.
[2] K. Al-Wahabi, R. Bari, F. Harary & D. Ullman, The edge-distinguishing chromatic number of paths and cycles, in Proceedings of Graph Theory in memory of G. A. Dirac (Sandbjerg, 1985) (eds. L. D. Andersen et al.), Annals of Discrete Mathematics, 41, (1989), pp. 17-22. [3] B. Auerbach & R. Laskar, Some coloring numbers for complete r-partite graphs, Journal of Combinatorial Theory, Series B, 21 (1976), 169-170.
[4] D. G. Beane, N. L. Biggs & B. J. Wilson, The growth rate of the harmonious chromatic number, Journal of Graph Theory, 13 (1989), 291-299.
[5] V. N. Bhat-Nayak & M. Shanthi, Achromatic number of Uz 1 Kn,, in Combinatorial Mathematics and Applications (Proceedings of the International Conference held in honor of Raj Chandra Bose, Calcutta, 1988) (eds. J. K. Ghosh et al.), Sankhya, Series A, 54, Special Issue, (1992), pp. 71-75. [6] V. N. Bhat-Nayak & M. Shanti, Achromatic numbers of a graph and its complement, Bulletin of the Bombay Mathematical Colloquium, 6 (1989), 9-14. [7] V. N. Bhave, On the pseudoachromatic number of a graph, Fundamenta Mathematicae, 102 (1979), 159-164. [8] H. L. Bodlaender, Achromatic number is NP-complete for cographs and interval graphs, Information Processing Letters, 31 (1989), 135-138.
[9] B. Bollobas, P. A. Catlin & P. Erdos, Hadwiger's conjecture is true for almost all graphs, European Journal of Combinatorics, 1 (1980), 195-199.
Harmonious Chromatic Number
41
[10] B. Bollobas, B. Reed & A. Thomason, An extremal function for the achromatic number, in Graph Structure Theory (Proceedings of the AMSIMS-SIAM Joint Summer Research Conference on Graph Minors, Seattle, 1991) (eds. N. Robertson & P. Seymour), Contemporary Mathematics, 147, American Mathematical Society, Providence, Rhode Island (1993), pp. 161-165. [11] F. Bories, Etude du nombre achromatique de certains graphes, in Colloque sur la Theorie des Graphes (Paris, 1974), Cahiers du Centre d'Etudes de Recherche Operationnelle, 17, (1975), pp. 155-171. [12] F. Bories, Sur quelques problemes de colorations completes de sommets et d'aretes de graphes et d'hypergraphes, These de 3eme cycle, Paris, 1975.
[13] F. Bories & J.-L. Jolivet, On complete colorings of graphs, in Recent Advances in Graph Theory, (Proceedings of Second Czechoslovak Symposium, Prague, 1974) (ed. M. Fiedler), Academia, Prague (1975), pp. 75-87. [14] A. Bouchet & R. Lopez-Bracho, Decomposition of a complete graph into trails of given lengths, Discrete Mathematics, 42 (1982), 145-152. [15] R. Brewster, Heuristics for computing the achromatic number of a graph, Manuscript, Victoria, B.C., 1987.
[16] R. Brewster & G. MacGillivray, Homomorphically full graphs, Discrete Applied Mathematics, 66 (1996), 23-31.
[17] B. E. Brunton, B. J. Wilson & T. S. Griggs, Graphs which have n/2minimal line-distinguishing colourings, Discrete Mathematics, 155 (1996), 19-26.
[18] F. Buckley & L. Superville, The ajointed number and graph homomorphism problems, in The Theory and Applications of Graphs (Proceedings
of Fourth International Conference on the Theory and Applications of Graphs, Kalamazoo, Michigan, 1980) (eds. G. Chartrand et al.), Wiley, New York (1981), pp. 149-158.
[19] N. Cairnie & K. J. Edwards, On the achromatic number of graphs, Manuscript, Dundee, 1996.
[20] G. Chartrand & J. Mitchem, Graphical theorems of the NordhausGaddum class, in Recent trends in Graph Theory (Proceedings of First New York City Graph Theory Conference, New York, 1970) (eds. M. Capobianco, J. B. Frechen & M. Krolik), Lecture Notes in Mathematics, 186, Springer, Berlin (1971), pp. 55-61.
Keith Edwards
42
[21] A. Chaudhary & S. Vishwanathan, Approximation algorithms for the achromatic number, in Proceedings of Eighth Annual ACM-SIAM Symposium on Discrete Algorithms (New Orleans, 1997), in press.
[22] N.-P. Chiang & H.-L. Fu, On the achromatic number of the cartesian product Gl x G2, Australasian Journal of Combinatorics, 6 (1992), 111117.
[23] C. A. Christen & S. M. Selkow, Some perfect coloring properties of graphs, Journal of Combinatorial Theory, Series B, 27 (1979), 49-59.
[24] R. J. Cichelli, Minimal perfect hash functions made simple, Communications of the Association for Computing Machinery, 23 (1980), 17-19.
[25] C. R. Cook & A. B. Evans, Graph folding, in Proceedings of 10th Southeastern Conference on Combinatorics, Graph Theory and Computing, Volume I (Boca Raton, 1979) (eds. F. Hoffman et al.), Congressus Numerantium, 23-24, (1979), pp. 305-314.
[26] K. J. Edwards, The harmonious chromatic number of almost all trees, Combinatorics, Probability and Computing, 4 (1995), 31-46. [27] K. J. Edwards, The harmonious chromatic number of bounded degree trees, Combinatorics, Probability and Computing, 5 (1996), 15-28. [28] K. J. Edwards, The harmonious chromatic number of bounded degree graphs, Journal of the London Mathematical Society, in press. [29] K. J. Edwards & C. J. H. McDiarmid, New upper bounds on harmonious colorings, Journal of Graph Theory, 18 (1994), 257-267.
[30] K. J. Edwards & C. J. H. McDiarmid, The complexity of harmonious colouring for trees, Discrete Applied Mathematics, 57 (1995), 133-144.
[31] M. Farber, G. Hahn, P. Hell & D. J. Miller, Concerning the achromatic number of graphs, Journal of Combinatorial Theory, Series B, 40 (1986), 21-39.
[32] B. Fawcett, On infinite full colourings of graphs, Canadian Journal of Mathematics, 30 (1978), 455-457.
[33] O. Frank, F. Harary & M. J. Plantholt, The line-distinguishing chromatic number of a graph, Ars Combinatoria, 14 (1982), 241-252. [34] M. R. Garey & D. S. Johnson, Computers and Intractability, W. H. Freeman, New York (1979).
Harmonious Chromatic Number
43
[35] D. P. Geller & S. T. Hedetniemi, A proof technique in graph theory, in Proof Techniques in Graph Theory (Proceedings of Second Ann Arbor Conference on Graph Theory, Ann Arbor, 1968) (ed. F. Harary), Academic Press, New York (1969), pp. 49-59.
[36] D. P. Geller & H. V. Kronk, Further results on the achromatic number, Fundamenta Mathematicae, 85 (1974), 285-290.
[37] D. P. Geller & S. Stahl, The chromatic number and other functions of the lexicographic product, Journal of Combinatorial Theory, Series B, 19 (1975), 87-95. [38] J. P. Georges, On the harmonious coloring of collections of graphs, Journal of Graph Theory, 20 (1995), 241-254.
[39] R. P. Gupta, Bounds on the chromatic and achromatic numbers of complementary graphs, in Recent Progress in Combinatorics (Proceedings of Third Waterloo Conference on Combinatorics, Waterloo, 1968) (ed. W. T. Tutte), Academic Press, New York (1969), pp. 229-235. [40] F. Harary & S. T. Hedetniemi, The achromatic number of a graph, Journal of Combinatorial Theory, 8 (1970), 154-161.
[41] F. Harary, S. T. Hedetniemi & G. Prins, An interpolation theorem for graphical homomorphisms, Portugaliae Mathematica, 26 (1967), 453-462.
[42] F. Harary & M. J. Plantholt, Graphs with the line-distinguishing chromatic number equal to the usual one, Utilitas Mathematica, 23 (1983), 201-207.
[43] W. R. Hare, S. T. Hedetniemi, R. Laskar & J. Pfaff, Complete coloring parameters of graphs, in Proceedings of 16th Southeastern International Conference on Combinatorics, Graph Theory and Computing (Boca
Raton, 1985) (eds. F. Hoffman et al.), Congressus Numerantium, 48, (1985), pp. 171-178.
[44] S. M. Hedetniemi, S. T. Hedetniemi & T. Beyer, A linear algorithm for the Grundy (coloring) number of a tree, in Proceedings of 13th Southeastern International Conference on Combinatorics, Graph Theory and Computing (Boca Raton, 1982) (eds. F. Hoffman et al.), Congressus Numerantium, 36, (1982), pp. 351-363. [45] P. Hell & D. J. Miller, On forbidden quotients and the achromatic number, in Proceedings of Fifth British Combinatorial Conference (Aberdeen, 1975) (eds. C. St.J. A. Nash-Williams & J. Sheehan), Congressus Numerantium, 15, (1976), pp. 283-292.
Keith Edwards
44
[46] P. Hell & D. J. Miller, Graph with given achromatic number, Discrete Mathematics, 16 (1976), 195-207. [47] P. Hell & D. J. Miller, Achromatic numbers and graph operations, Discrete Mathematics, 108 (1992), 297-305.
[48] J. E. Hopcroft & M. S. Krishnamoorthy, On the harmonious coloring of graphs, SIAM Journal on Algebraic and Discrete Methods, 4 (1983), 306-311. x [49] M. Horiiak & J. Puntigan, On the achromatic number of KK, in Graphs and other Combinatorial Topics, Proceedings of
Third Czechoslovak Symposium on Graph Theory, Prague, 1982) (ed. M. Fiedler), Teubner- Texte zur Mathematik, 59, Teubner, Leipzig (1983), pp. 118-123.
[50] F. Hughes, On the achromatic number of graphs, M.Sc. Thesis, University of Victoria, B.C., 1994. [51] F. Hughes & G. MacGillivray, The achromatic number of graphs: a survey
and some new results, Bulletin of the Institute of Combinatorics and its Applications, in press.
[52] T. R. Jensen & B. Toft, Graph Coloring Problems, Wiley, New York (1995).
[53] J.-L. Jolivet, Graphes parfaits pour une propriete P, in Colloque sur la Theorie des Graphes (Paris, 1974), Cahiers du Centre d'Etudes de Recherche Operationnelle, 17, (1975), pp. 253-256.
[54] J. Kelly, Difference systems, graph designs, and coloring problems, Journal of Combinatorial Theory, Series B, 30 (1981), 144-165. [55] I. Krasikov & Y. Roditty, Bounds for the harmonious chromatic number of a graph, Journal of Graph Theory, 18 (1994), 205-209.
[56] A. Kundrik, The harmonious chromatic number of a graph, in Proceedings of Fourth Czechoslovakian Symposium on Combinatorics, Graphs and Complexity (Prachatice, 1990) (eds. J. Nesetril & M. Fiedler), Annals of Discrete Mathematics, 51, (1992), pp. 167-170.
[57] S.-M. Lee & J. Mitchem, An upper bound for the harmonious chromatic number of a graph, Journal of Graph Theory, 11 (1987), 565-567.
[58] R. Lopez-Bracho, Etudes du nombre achromatique des etoiles, Thesis, University de Paris-Sud, 1981. [59] R. Lopez-Bracho, Le nombre achromatique d'une ytoile, Ars Combinatoria, 18 (1984), 187-194.
Harmonious Chromatic Number
45
[60] Zhikang Lu, On an upper bound for the harmonious chromatic number of a graph, Journal of Graph Theory, 15 (1991), 345-347. [61] Zhikang Lu, The harmonious chromatic number of a complete binary and trinary tree, Discrete Mathematics, 118 (1993), 165-172.
[62] Zhikang Lu, Estimates of the harmonious chromatic numbers of some classes of graphs (Chinese), Journal of Systems Science and Mathematical Sciences, 13 (1993), 218-223.
[63] Zhikang Lu, The harmonious chromatic number of a complete 4-ary tree, Journal of Mathematical Research and Exposition, 15 (1995), 51-56.
[64] A. Mate, A lower estimate for the achromatic number of irreducible graphs, Discrete Mathematics, 33 (1981), 171-183.
[65] C. J. H. McDiarmid, Colourings random graphs badly, in Graph Theory and Combinatorics (Proceedings of a Conference, Milton Keynes, 1978) (ed. R. J. Wilson), Pitman Research Notes in Mathematics, 34, Pitman, San Francisco (1979), pp. 76-86. [66] C. J. H. McDiarmid, Achromatic numbers of random graphs, Mathematical Proceedings of the Cambridge Philosophical Society, 92 (1982), 21-28. [67] C. J. H. McDiarmid & Luo Xinhua, Upper bounds for harmonious colorings, Journal of Graph Theory, 15 (1991), 629-636.
[68] F. Milazzo & V. Vacirca, On the achromatic number of permutation graphs, in Proceedings of First Catania International Combinatorial Con-
ference on Graphs, Steiner Systems and their Applications, Volume 2 (Catania, 1986) (ed. M. Gionfriddo), Ars Combinatoria, 24B, (1987), pp. 71-76. [69] F. Milazzo & V. Vacirca, On the achromatic number of G x Km, in Proceedings of First Catania International Combinatorial Conference on Graphs,
Steiner Systems and their Applications, Volume 2 (Catania, 1986) (ed. M. Gionfriddo), Ars Combinatoria, 24B, (1987), pp. 173-177. [70] Z. Miller, Extremal regular graphs for the achromatic number, Discrete Mathematics, 40 (1982), 235-253.
[71] Z. Miller & D. Pritikin, The harmonious coloring number of a graph, in Proceedings of 250th Anniversary Conference on Graph Theory (Fort Wayne, Indiana, 1986) (eds. K. S. Bagga et al.), Congressus Numerantium, 63, (1988), pp. 213-228.
[72] Z. Miller & D. Pritikin, The harmonious coloring number of a graph, Discrete Mathematics, 93 (1991), 211-228.
Keith Edwards
46
[73] J. Mitchem, On the harmonious chromatic number of a graph, Discrete Mathematics, 74 (1989), 151-157. [74] C. St.J. A. Nash-Williams, Detachments of graphs and generalised Euler trails, in Surveys in Combinatorics 1985 (Invited papers for Tenth British Combinatorial Conference) (ed. I. Anderson), Cambridge University Press, Cambridge (1985), pp. 137-151. [75] N. Pippenger & J. Spencer, Asymptotic behavior of the chromatic index for hypergraphs, Journal of Combinatorial Theory, Series A, 51 (1989), 24-42. [76] F. Regonati & N. Zagaglia Salvi, Some constructions of )A-minimal graphs, Czechoslovak Mathematical Journal, 44(119) (1994), 315-323.
[77] F. Regonati & N. Zagaglia Salvi, Minimal semicomplete graphs, in Proceedings of International Conference on Graphs and Hypergraphs (Var-
enna, 1991) (eds. O. D'Antona, M. Gionfriddo & N. Zagaglia Salvi), Journal of Combinatorics, Information and System Sciences, 19, Forum for Interdisciplinary Mathematics, Delhi (1994), pp. 63-73. [78] E. Sampathkumar & V. N. Shave, Partition graphs and coloring numbers of a graph, Discrete Mathematics, 16 (1976), 57-60.
[79] A. J. Schwenk, Almost all trees are cospectral, in New Directions in the Theory of Graphs (Proceedings of Third Ann Arbor Conference on Graph Theory, Ann Arbor, 1971) (ed. F. Harary), Academic Press, New York (1973), pp. 275-307.
[80] B. J. Wilson, Line-distinguishing and harmonious colourings, in Graph Colourings (Proceedings of a Conference on Graph Colourings, Milton Keynes, 1988) (eds. R. Nelson & R. J. Wilson), Pitman Research Notes in Mathematics, 218, Longman Scientific & Technical, Essex (1990), pp. 115133.
[81] B. J. Wilson, Minimal line distinguishing colourings in graphs, in Proceedings of Combinatorics '90 (Gaeta, 1990) (eds. A. Barlotti et al.), Annals of Discrete Mathematics, 52, (1992), pp. 549-558.
[82] R. M. Wilson, Decomposition of complete graphs into subgraphs isomorphic to a given graph, in Proceedings of Fifth British Combinatorial Conference (Aberdeen, 1975) (eds. C. St.J. A. Nash-Williams & J. Sheehan), Congressus Numerantium, 15, (1976), pp. 647-659. [83] Shaoji Xu, Relations between parameters of a graph, Discrete Mathematics, 89 (1991), 65-88.
Harmonious Chromatic Number
47
[84] M. Yannakakis & F. Gavril, Edge dominating sets in graphs, SIAM Journal on Applied Mathematics, 38 (1980), 364-372.
[85] N. Zagaglia Salvi, A note on the line-distinguishing chromatic number and the chromatic index of a graph, Journal of Graph Theory, 17 (1993), 589-591.
Department of Mathematics and Computer Science University of Dundee Dundee DD1 4HN U.K.
[email protected] Computer Construction of Block Designs Clement Lam Summary This paper uses an extended example to illustrate how to put together a BDX program to construct block designs fixed by an automorphism, given its orbit matrix. It shows how to specify the parameters and the structural information of the designs. It discusses the symmetry group of the problem and isomorph rejection. It explains how to choose a good order of generation to minimize the size of the search. It also shows how to estimate the size of a search and how to partition the problem into subproblems which can be searched in parallel on several computers. 1
Introduction In a recent paper that I co-authored [1], we wrote:
"For each orbit matrix, we used the BDX program to try out all possible circulant matrices with the correct row sum." Here, I would like to expand on this sentence, not so much to bore you with details, but to use it as an example to explain how to use the BDX program. This paper is intended as a companion to the BDX reference Guide [7], which
is a dry document outlining the syntax and meaning of each of the BDX commands. In this paper, I shall illustrate how the commands can be put together to solve a real problem. Let me first state what problem we are trying to solve. We want to find all quasi-symmetric 2-(28,12,11) designs with intersection numbers 4 and 6, which are fixed by an automorphism of order 7 without fixed points or fixed blocks.
I assume you know what 2-(v, k, A) designs are. I will introduce the other terminology as we need it. There are two other parameters associated with a 2-(v, k, A) design D, namely b, the number of blocks, and r, the number of replications. In fact, a 2-(v, k, A) design is often referred to as a (v, b, r, k, A) design. For the 2-(28,12,11) design, the corresponding values are b = 63 and r = 27.
Let P1, ..., P be a labelling of the points of D and let B1, ..., Bb be a labelling of the blocks of D. We define a matrix A = (ate) by a2j = 1 if Pi is on Bj and aia = 0 otherwise. Thus A is a v x b matrix with each entry either 1 or 0. We call A an incidence matrix for D. Result 1.1 If A is an incidence matrix for a 2-(v, k, A) design, then A' is the transpose of A, I, is the v x v identity matrix, J is the v x v matrix all of whose entries are +1 and n = r - A. 49
Clement Lam
50
BDX is a computer program that is designed to explore how best to enumerate 2-designs. It is built to be easily adaptable to designs of different parameters. It treats all designs as incidence matrices. As a consequence, we tend to treat the words rows and points, as well as the words columns and blocks, as synonyms. BDX takes a sequence of commands as input, executes them and produces
output. For example, if you want to enumerate all 2-(28,12,11) designs, you can input: 28 rows of 27; 63 cols of 12; lambda = 11; try row [1:28, 1:631;
The first three lines define a 28 x 63 (0,1)-matrix with 27 ones in each row, 12 ones in each column and in which the inner product of any pair of distinct rows is 11. This defines an incidence matrix of a 2-(28,12,11) design. The last line instructs BDX to find all possible ways of putting 1's and 0's into this matrix subject to the constraints given. I would not suggest you execute this input. The program would run for an extremely long time and find lots of solutions, but will not tell you what the solutions are. If it ever finishes, it will tell you how many solutions it found; but no one is going to live long enough to see it finish. In the remainder of this paper, Section 2 shows how to formulate the problem for BDX and Section 3 demonstrates how to solve the problem using BDX. 2
Formulating the problem for BDX
A keen reader may observe that we are not trying to find all 2-(28,12,11) designs, but rather a restricted subset. Let us first motivate the definition of the quasi-symmetric property. While the inner products of pairs of distinct rows of the incidence matrix of a block design are always A, the inner product of pairs of columns are not necessarily constant. These inner products of distinct columns are the intersection numbers of the block design.
A symmetric design is one where v = b and, consequently, r = k. A symmetric design has a unique intersection number A [5]. A quasi-symmetric design is one with only two intersection numbers [6]. In our case, we are looking for an incidence matrix where the inner products of pairs of distinct columns are either 4 or 6. How do we specify the quasi-symmetric property to BDX? In the literature, A'A is often referred to as the block intersection matrix S [2]. So, in BDX,
S refers to this block intersection matrix. The BDX command to define a quasi-symmetric design with intersection numbers 4 and 6 is:
Computer Construction of Block Designs
51
init S = {4 & 6};
It initializes the set of possible values for the off-diagonal entries of S to {4,6}. An automorphism of a block design is an incidence preserving relabelling
of its points and blocks. Let us restate it in terms of an incidence matrix A. Let 7r be a permutation of its columns and let p be a permutation of its rows. The pair of row and column permutations (p, 7r) is an automorphism of A if aij = ap(i),'Iro) for all rows i and columns j. In our problem, we are looking for all 2-(28,12,11) designs which are fixed by an automorphism of order 7 with no fixed points or blocks. This automorph-
ism partitions the points into 4 orbits and the blocks into 9 orbits, where all the orbits are of size 7. By reordering the rows and columns, we can have a nice presentation of the structures of both the incidence matrix and the row and column permutations. In particular, we choose the image of row 1 under p as row 2, the image of row 2 as row 3, and so on. Because the orbits are of size 7, the image of row 7 is row 1. Thus, any row can be row 8, and we can then continue to use the image of row 8 as row 9, etc. After applying this process to all the rows, we can assume that p(2)
_
i+1 if 7 A' i, and i - 6 otherwise.
Similarly, applying the process to all of the columns, 7r
(j)=I +1 if7%j,and j-6
otherwise.
In BDX, permutations are input in the image form, which simply means that, for a row permutation, we list all the values of p(i), i = 1.... , v, and, for a column permutation, we list 7r(j), j = 1, . . . , b. Thus, we specify this assumed automorphism information to BDX by: assume row perm = 2 3 4 5 6 7 1 23 24 25 26 27 28 22; assume col perm = 2
3
4
5
6
7
1
23 24 25 26 27 28 22 44 45 46 47 48 49 43
9 10 11 12 13 14
8
9 10 11 12 13 14 8 30 31 32 33 34 35 29 51 52 53 54 55 56 50
16 17 18 19 20 21 15
16 17 18 19 20 21 15 37 38 39 40 41 42 36 58 59 60 61 62 63 57;
BDX uses the assumed automorphism to derive the implications of placing a value into the incidence matrix. For example, if a1,1 is 0, then a2,2 = ap(l),,,(1)
is also 0. More generally, if aij is x, then ap(i),,(A) is also x. Thus, with the assumption of this automorphism, the incidence matrix for the 2-(28,12,11) design can be partitioned into 4 x 9 blocks of 7 x 7 circulant matrices. For later reference, let us label these circulant matrices as Ai,j, i = 1, ... , 4 and
j = 1,...,9.
Clement Lam
52 0
3
3
3
3
3
4
4
4
4
1
3
3
3
5
2
2
4
4
3
3
3
5
1
2
4
2
4
5
3
3
1
3
4
2
2
Figure 1: The orbit matrix for case 26
A circulant matrix has constant row and column sums. The collection of row and column sums of all the circulant matrices gives a set of structural information about the incidence matrix. The orbit matrix or tactical decomposition of A with respect to the assumed automorphism is a 4 x 9 matrix C = (ci,j), where ci,j is the column sum of the (i,j)-th circulant matrix Ai,j. The entries in an orbit matrix satisfy two relations [3], which in our case trans-
late to c
9
k jk
k= 1
=
J 77 ifi j,1 H is a group isomorphism, cp: 1 E is a bijection and for all g E G, a E Q, we have (a9)W = (acp)i9lf. As abstract groups Gp is a quotient group of G, and we view the permuta-
tion group Gp on P as a quotient of the the permutation group G on 0. If we choose P such that Gp is maximal in G for some, and hence all, p E P, we see that Gp must be primitive on P. Thus every transitive permutation group has at least one primitive quotient. Similarly if P is a G-normal partition then the permutation group G' will be called a normal quotient of G. By choosing a normal subgroup N to be maximal by inclusion such that N is intransitive on Q, we can ensure that the normal quotient GIN is quasiprimitive. Thus every transitive permutation group has at least one quasiprimitive normal quotient. One reason for making the latter definition is that there is a useful analogous concept for graphs. For a graph F = (V, E) with vertex set V and edge
set E, if P is a partition of V then the quotient graph of r relative to P is defined to be the graph Fp = (P, Ep) where {p, p'} E Ep if and only if there
Finite quasiprimitive graphs
67
exist x E p and x' E p' such that {x, x'} E E. Quotient graphs inherit some of the properties of the original graphs. For example, quotients of connected graphs are connected. Moreover if the partition P is invariant under some subgroup G of the automorphism group Aut(F), then Fp admits G' as a subgroup of automorphisms. In particular, if G < Aut(F) and G is arc-transitive on r, then for each G-invariant partition P of V, G' is arc-transitive on Fp. If P is a G-normal partition of V, say P is the set of N-orbits in V with N a normal subgroup of G, then we say that Fp is a G-normal quotient of r, and
write rN = rp. Note that, if F is connected and G is arc-transitive, then for any nontrivial G-invariant partition P, each edge of r joins vertices in distinct parts of P. Thus each p E P is an independent subset of V and the subgraph induced on p is an empty graph. Moreover, since in this case G is transitive on unordered pairs of adjacent parts {p, p'} in Vp, the subgraph of r induced on p U p' is independent of the pair {p, p'} E Ep. In the special case where the edges of this subgraph form a complete matching between p and p', that is where each vertex of p is joined to exactly one vertex of p', we say that r is a cover of Fp. (This terminology comes from topology with the graph r playing the role of a covering space and the parts of the partition P corresponding to the fibres.) 3
Primitive graph quotients
If G is imprimitive on the vertex set V of F then, as explained above, we may choose a G-invariant partition P of V such that Gp is primitive on the quotient graph Vp, and a G-normal partition PN such that GPN is quasiprimitive on the normal quotient FN. Of course if r is bipartite then the primitive or quasiprimitive quotient may be simply the complete graph K2 on two vertices. However for non-bipartite graphs IT, depending on the combinatorial relationships between r and its quotients, or normal quotients, a great deal may be learned about r from the set of its primitive quotients, or quasiprimitive normal quotients respectively. In this and the following sections we shall explore these relationships further for several families of finite arc-transitive graphs which have group theoretic defining properties. Whether the primitive or the quasiprimitive quotients
are the appropriate ones to study tends to depend on whether the defining property of the family is essentially a global property or a local property. By a local property we mean a condition on the neighbourhood of a vertex (or sometimes on the vertices up to some fixed distance from a given vertex). Families with a global group theoretic defining property are sometimes closed under the formation of nearly all graph quotients relative to partitions invari-
ant under the relevant group, and for these families it is useful to study the primitive quotients of graphs in the family. On the other hand, some families of arc-transitive graphs with a local group theoretic defining property are closed under taking normal quotients, but are not closed under taking arbitrary quo-
68
Cheryl E. Praeger
tients relative to partitions invariant under the given group. For these families it is the quasiprimitive normal quotients (that is the quasiprimitive graphs in the family) which give better information about typical graphs in the family. We give some examples below. For a graph F = (V, E), the distance d(a, /3) between two vertices a and 0
is the length of the shortest path between them. In particular d(a, a) = 0 for each a E V. If r is finite and connected then the maximum d(F) of the distances between pairs of vertices of F is called the diameter of F. For example, if IVI = n > 1, then d(F) = 1 if and only if I is the complete graph Kn. For i = 0, ... , d(F), we set Fi := {(a, 6) : d(a, /3) = i}, and Fi(a) := {,Q : i}. Thus 171 (a) is the set of neighbours of a in r, and we sometimes write this set
simply as F(a). Example 3.1 A graph r is said to be distance transitive (or G-distance transitive where G < Aut(F)) if, for each i = 0, . . . , d(F), Aut(F) (respectively G) is transitive on IT,. If r = (V, E) is a G-distance transitive graph, and P is a nontrivial G-invariant partition of V, then it was shown by D. H. Smith [31] (or see [5, Theorem 4.2.1, and the following remarks]) that either r is bipartite, P is the bipartition of V, and Fp = K2, or F1, is GP-distance transitive and r is a cover of FP. Thus non-bipartite finite distance transitive graphs are covers of vertex-primitive distance transitive graphs. In the case of a finite bipartite G-distance transitive graph r with bipartition P, Smith also showed that the distance-two graphs defined on the two parts of P are isomorphic distance transitive graphs, and are not themselves bipartite provided r has valency at least 3. (For p E P, the distance-two graph on p is the graph with vertex set p and edges {{a, /3} : E p, d(a, Q) = 2}.) Thus, in a strong sense, understanding the finite primitive distance transitive graphs is the key to understanding the whole family of finite distance transitive graphs. There are many infinite series of finite primitive distance transitive graphs. One which is easy to describe is the family of Hamming
graphs H(d, m) where d > 2, m > 3. The vertex set of H(d, m) is Z and vertices x = (X1, ... , xd) and y = ( Y 1 ,- .. , Yd) are joined by an edge if and only
if x - y has weight 1, that is xi = yi for all except one value of i = 1, .
. . ,
d.
The diameter of H(d, m) is d and its automorphism group is S,n wr Sd. It turns out that the complement H(2, m) of H(2, m) is also distance transitive for all
m > 3. In [28] a study was begun of finite primitive distance transitive graphs guided by the O'Nan-Scott Theorem [17]. It was shown there that, if r = (V, E) is a G-distance transitive graph with G primitive on vertices, then either F is H(d, m) or H(2, m) for some d > 2, m > 3, or G is a primitive group on V of primitive type HA or AS. These types of primitive permutation groups are
defined as follows. A primitive permutation group G of type HA on Sl is a group of affine transformations of a finite vector space V, and is of the form G = NH where N is the group of translations of V, H is an irreducible group
Finite quasiprimitive graphs
69
of nonsingular linear transformations of V, and St = V with the natural action. A primitive permutation group G on S2 has type AS if G is an almost simple group, that is if T < G < Aut(T) for some finite nonabelian simple group T. Much progress has been made on completing the classification of finite primitive distance transitive graphs, and a good account of this can be found in [12]. Completing this classification is feasible, and it will rely on the finite simple group classification. However it is not yet complete.
Problem 3.2 Complete the classification of the finite primitive distance transitive graphs.
Example 3.3 Let s be a positive integer. An s-arc in a graph r = (V, E) is a sequence a = (ao, all ... , as) of vertices such that Jai-1, ai} E E for i = 1, . . . , s. So in particular a 1-arc is simply an arc of F. The graph r is said to be (G, s)-arc transitive if G < Aut(F) and G is transitive on the set of s-arcs of F. If such a subgroup G exists, we also say simply that F is s-arc transitive. For a vertex-transitive graph, the property of s-arc transitivity involves the vertices at distance up to Is/2l from a given vertex and so may be regarded as a kind of generalised group theoretic local condition on F. (To see this consider all the s-arcs a with a given vertex as aL512j.) Thus it is the cases s = 1, 2 which provide genuinely local group theoretic conditions, that is conditions on the neighbourhood of a vertex. To be precise, if G is vertex-transitive on r
then, for s = 1, 2, F is (G, s)-arc transitive if and only if Ga is s-transitive on F(a). (A permutation group is s-transitive if it is transitive on ordered s-tuples of distinct points.) We have already observed that for s = 1, if I' is (G, s)-arc transitive and if P is a nontrivial G-invariant partition of V, then FP is (GP, s)-arc transitive. However this implication is not true if s = 2. The situation may be illustrated effectively in the case of (G, 2)-arc transitive graphs with G a Suzuki simple group Sz(q), with q = 22m+1 > 8. This case may be typical of many families of (G, 2)-arc transitive graphs with almost simple automorphism groups G. Let q = r2/2, and let t be a divisor of 2m + 1, t > 1. Then 5 divides q + Sr + 1, where 5 is 1 or -1. There are (Sz(q), 2)-arc transitive graphs r(5) and F(2t) of valencies 5 and 2t respectively. These graphs have Sz(q)-vertex primitive quotients F(5)P and F(2t)y such that P has parts of size (q + Sr + 1)/5 and q2(q - 1)/ (2t(2t - 1)) respectively. Also between each pair of adjacent parts of P there is a single edge or (q - 1)/(2t - 1) edges respectively, so that r(5) and F(2t) are far from being covers of these quotients. Moreover F(5)p has valency q + Sr + 1 and is not 2-arc transitive, while F(2t)y is the complete graph K92+1 on which (by "chance") Sz(q) acts 2-arc transitively. (These are some of the 2-arc transitive graphs constructed by Xin Gui Fang in his PhD thesis [7], see also [8].)
Thus the property of being 2-arc transitive is not in general inherited by primitive quotients, so a study of finite primitive 2-arc transitive graphs will
Cheryl E. Praeger
70
not yield significant information about typical finite 2-arc transitive graphs. In contrast to this we shall show in the next section that each finite non-bipartite 2-arc transitive graph P is a cover of a quasiprimitive 2-arc transitive graph.
4
Normal quotients of locally primitive and locally quasiprimitive graphs
In this section we study several families of vertex transitive graphs which are subfamilies of the family .F1_arc of finite 1-arc transitive graphs and which contain the family .F2_arc of finite 2-arc transitive, vertex transitive graphs. For
a group theoretic property Q, we shall say that a graph r is G-locally Q (or simply locally Q) if the action induced by the vertex stabilizer Ga on r(a) has property Q. Thus, within the class of finite vertex transitive graphs, .F2-arc is the family of locally 2-transitive graphs, and F1-arc is the family of locally transitive graphs. Since all 2-transitive permutation groups are primitive, T2_arc is contained in the family .i-prim of finite vertex transitive, locally primitive graphs. Similarly, since all primitive permutation groups are quasiprimitive, .F1_prim is contained in the family .F1_gprim of finite vertex transitive, locally quasiprimitive graphs, and of course F,-,c contains .F'1_gprim . Thus T2-arc C -.l-prim C J71-gprim C .F1-arc
(1)
We shall investigate normal quotients of graphs in these families. To describe the structure of these normal quotients we need some more notation. A permutation group G on a set Sl is said to be semiregular on SZ if the only element of G which fixes a point of Sl is the identity element 1c. Let P be a partition
of the vertex set of a graph r = (V, E). Then r will be called a multicover of Fy if, for each {p, p'} E EP, each vertex of p is joined to at least one vertex of p'. The multicovers we describe below are all uniform in the sense that, IF(a) np'I is independent of the choice of the vertex a in p. Multicovers were called pseudocovers in [24], but the term multicover seems a better name for them. The theorem below was proved in [24, Section 1].
Theorem 4.1 Let r = (V, E) be a connected (G, 1)-arc transitive graph of valency v, and let N be a normal subgroup of G. Then one of the following holds.
(a) N is transitive on V; or (b) F is bipartite and the N-orbits in V are the two parts of the bipartition; or
(c) FN = (PN, EN) is a connected (G'N,1)-arc transitive graph of valency v/k where, for each {p, p'} E EN and each a E p, IF(a) fl p'l = k, and I' is a multicover of FN. Moreover,
Finite quasiprimitive graphs
71
(i) if in addition r is G-locally quasiprimitive then N is semiregular on V and FN is GIN-locally quasiprimitive (and the permutation group G. acts faithfully on the partition P(a) defined by P(a) {F(a) fl p' : {p, p'} E EN});
(ii) if F is G-locally primitive then r is a cover of FN (that is k = 1) and FN is GIN-locally primitive (and the permutation groups Gr(') and GPN(p) are permutationally equivalent); (iii) in particular if r is (G, 2)-arc transitive then r is a cover of FN and FN is (GPN, 2)-arc transitive.
Thus, if we restrict ourselves to the non-bipartite members r of one of the families in (1) we see that all of the normal quotients of F lie in the same family as F does, and each non-bipartite graph in each of these families has at least one quasiprimitive normal quotient. Moreover each non-bipartite graph in .Fi-prim or .F2-arc is a cover of a quasiprimitive graph in .9-prim or ..T2_arc respectively. Thus the quasiprimitive graphs in each family in (1) are important for understanding the structure of typical non-bipartite graphs in the family. However to test the power of this observation, and ultimately to gain some more concrete knowledge of these families of graphs, we need some detailed information about the structure of finite quasiprimitive permutation groups, similar to the information about finite primitive permutation groups
given by the O'Nan-Scott Theorem. At the time when Theorem 4.1 was proved no information of this nature was available in the literature. Thus we first needed to analyse the structure of finite quasiprimitive permutation groups. This analysis was undertaken in [26] and will be discussed in the next section. Then, in Section 6 we will illustrate, with the family F'2-arc, how this structure theory may be applied to help understand the structure of quasiprimitive graphs in certain families of graphs.
It should be emphasised that, of the four families in (1), it is only the family F1-arc that is closed under taking arbitrary primitive quotients. The 5-valent (Sz(q), 2)-arc transitive graph F(5) described in Example 3.3 exhibits
properties which may be typical of many other (G, 2)-arc transitive graphs with G almost simple. The primitive (Sz(q),1)-arc transitive quotient F(5)P of valency q + Jr + 1 described there is not even Sz(q)-locally quasiprimitive if
q>8. Appropriate means for understanding the bipartite examples in these families are more difficult to find. In the case of bipartite finite distance transitive graphs the distance-two graph induced on a part of the bipartition turns out to be again distance transitive. However each of the families in (1) has a local group theoretic defining property, and consequently the situation is rather different from that for finite distance transitive graphs. There is no natural way to obtain a graph in the relevant family induced on a part of the bipartition
of a bipartite graph in one of the families in (1). An approach to studying
Cheryl E. Praeger
72
the structure of finite bipartite 2-arc transitive graphs as covers of smaller graphs in J 2_arc was made in [27]. The relevant quotients in this case are those for which every nontrivial normal subgroup of the group in question is either vertex-transitive or has as vertex orbits the two parts of the bipartition. The action induced on a bipartite half of such a quotient is not necessarily quasiprimitive. Nevertheless the structure theorem for quasiprimitive groups described in the next section is crucial in analysing the possible structures of such quotients: the results are not as nice as those for the non-bipartite case. 5
Finite quasiprimitive permutation groups
The class of finite quasiprimitive permutation groups may be described in a fashion very similar to the description given by the O'Nan-Scott Theorem [17, 30] for finite primitive permutation groups. This provides essentially an identification of several types of finite quasiprimitive groups such that for each type we have additional information about either the abstract group theoretical structure, or the nature of the action, or both. The quasiprimitive types in most cases are similar to the primitive types from the O'Nan-Scott Theorem. The ordering and the presentation of the various types, and the amount of subdivision of the types preferred by those who wish to use this classification vary according to the requirements of the different applications. For some requirements the nature of the socle or the minimal normal subgroups is most important, while (for example) for others it is the nature of the suborbits, or the existence of regular subgroups that is needed. (The socle soc(G) of a group G is the product of the minimal normal subgroups of G. A permutation group G on S2 is said to be regular on SZ if G is both transitive and semiregular on 0.) For this exposition I have chosen a subdivision into types which was suggested by Laci Kovacs in 1985 for the types of primitive groups in the O'NanScott Theorem. It is a little finer than the subdivision used in the original paper on the structure of quasiprimitive groups [26], which mirrored the subdivision given in [17] for primitive groups. The subdivision given here was used in [25] for primitive groups and again in [3] for quasiprimitive groups; and moreover in the latter paper a slight refinement of it was introduced. We shall define eight types of finite quasiprimitive permutation groups. The main theorem in [26] states that every finite quasiprimitive permutation group belongs to exactly one of these types. Let G be a finite quasiprimitive permutation group on S2 and let a E Q. Then G has at most two minimal normal subgroups (see the first portion of [26, Section 3]), and if there are two minimal normal subgroups then they are isomorphic and each is equal to the centraliser of the other. In particular if G has an abelian minimal normal subgroup N then N is the unique minimal normal subgroup of G, that is N is the socle of G. Further N is elementary abelian and is regular on S2, and G is the semidirect product G = NG,,, for a E Q. Thus we may identify fZ with
Finite quasiprimitive graphs
73
N = Z P', which may be viewed as a d-dimensional vector space over a field of prime order p, and choosing a as the zero vector we then have that Ga is an irreducible subgroup of nonsingular linear transformations of N. This is our first quasiprimitive type, and is named HA since G is contained in the Holomorph of the Abelian group N. The holomorph Hol(N) of a group N is the semidirect product N. Aut(N), where Aut(N) acts naturally on the normal subgroup N.
HA SZ = 7Ld for a prime p and positive integer d and G is the semidirect product G = N.G0, a subgroup of the affine group AGL(d,p) on S2, where N is the group of translations and Go is an irreducible subgroup of GL(d,p). Note that, for this quasiprimitive type, the requirements for quasiprimitivity and primitivity coincide so that a quasiprimitive permutation group of type HA is primitive. For all other types each minimal normal subgroup N of G is nonabelian, and hence N = Tl x ... x Tk for some positive integer k where each Ti T, a nonabelian simple group. If there is a second minimal normal subgroup M N, then M ^_' N, both M and N are regular on SZ, and soc(G) = N x M. In this case G is contained in the holomorph of N, and the second minimal normal subgroup M = {cp;'x : x E N}, where cps is the inner automorphism of N induced by x. This case is subdivided into two quasiprimitive types, namely the cases where k = 1 and k > 1. In the former case the type is named HS since G is contained in the Holomorph of the Simple group T. The case where k > 1 is named HC since G is contained in the Holomorph of a Compound group N. (The word "compound" is used here because G is also a subgroup of a group of compound diagonal type, a type which will be described
later.) Again for these types HS and HC the conditions for quasiprimitivity and primitivity coincide and each quasiprimitive permutation group of type HS or HC is primitive. For a group N, and for x E N and a E Aut(N), we {cp., : x E N}, denote the image of x under o, by x°. Also we set Inn(N) the inner automorphism group of N.
G < Hol(T) _ T. Aut(T), where for x E S2, t E T and a E Aut(T), to: x -r x°t°. If
HS Q = T, a finite nonabelian simple group, and T. Inn(T) a = 1T E SZ, then Inn(T) < Ga < Aut(T).
HC SI = T' = N, where k > 1 and T is a finite nonabelian simple group,
and N. Inn(N) < G < Hol(N) = N. Aut(N). As for the type HS, for x, n E N and a E Aut(N), nor: x H x°n°. If a = 1N E S2, then Inn(N) < Ga < Aut(N) and Ga acts transitively by conjugation on the simple direct factors of N.
There are five further quasiprimitive types and for each type there is a corresponding type of primitive group. However, unlike the types HA, HS,
Cheryl E. Praeger
74
and HC, each of these five types contains imprimitive quasiprimitive permutation groups. Moreover, for each of these quasiprimitive types N = soc(G) is the unique minimal normal subgroup of G, and for the first four of them the parallel with the corresponding primitive type is very close. These four types
are the type AS where G is Almost Simple, type SD where G has a Simple Diagonal action on S2, type CD where G has a Compound Diagonal action on Q, and type TW where G is a Twisted Wreath product with regular socle.
AS T < G < Aut(T), where T is a finite nonabelian simple group, and G = TGa with T Ga. In this type we may have Ga = {1G}, that is N = T may be regular on Q. For G to be primitive of type AS the stabilizer Ga must be maximal in G, and in particular Ga # {1G}. The next type is the simple diagonal type SD. A quasiprimitive permutation group of type SD is a subgroup of the group
W = {(aii... , ak)7r : ai E Aut(T), 7r E Sk, ai - aj (mod Inn(T)) for all i, j} where it-1(a1 i ... , ak)ir = (a1,r-1, ... , ak,r-1) and k > 1. The socle of W is
soc(W) = {(t1, ... , tk) ti E Inn(T)}, and a primitive action of W on
Tk-1
(which we identify with Inn(T)k-1) is defined by (ai, ... , ak): (t1, ... , tk_1) H (a, lt1a1i... , ak 1tk-lak-1), 7r: ( t 1 ,
and
-'-
. .. , tk_1) H (t-1
for (a1,.. . , ak)lr E W and (t1, ... , tk_1) E Tk-1, where tk = 1T. Thus for a = (1T, ... ,1T) E Tk-1, Wa = A x Sk where A = {(a, ... , a) : a E Aut(T)}.
SD Q = Tk-1, where k > 1 and T is a finite nonabelian simple group, N = soc(W) < G < W with the action defined above, and G acts transitively by conjugation on the k simple direct factors of N. The name Simple Diagonal comes from the fact that Na is the full diagonal subgroup {(t, . . . , t) : t E Inn(T)} of N. For the corresponding primitive type SD, Ga must act primitively on the simple direct factors of N. For the compound diagonal type CD the group G preserves a product structure on S2, that is SZ = Al for some l > 2 and G < Sym(A) wr S` where, for 8 = (81 i ... , 81) E S2 and (a1i . . . , a,) 7r E Sym(A) wr Sl, (a1, ... , at): 8 -+ (811, ... , 8`'),
and
ir: 8
Moreover the subgroup of Sym(A) involved is a quasiprimitive group of type SD, hence the name Compound Diagonal.
CD 1 = A' and N = Tk < G < H wr Sl < Sym(A) wr SS, for some divisor 1 of k, where l > 2 and kll > 2, T is a finite nonabelian simple group, H < Sym(A), soc(H) = Tk1`, and H is quasiprimitive on A of type SD; also G acts transitively by conjugation on the simple direct factors of N.
Finite quasiprimitive graphs
75
For the primitive type CD the subgroup H induced on A must be primitive
of type SD. The next type is the twisted wreath type TW. The original definition of a twisted wreath product was given by B. H. Neumann in [22]. The exposition here follows that in Suzuki's book [32, p. 269] and is the same as in [17]. Its use was suggested by Laci Kovacs. The core of a subgroup H of a group G is coreG(H) := ngEG H9. The twisted wreath product T twr,, P of groups T and P relative to cp may be defined as follows. Let P have a transitive action on {1, . . . , k} and let Q be the stabilizer of the point 1 in this action. Suppose that there is a homomorphism c': Q -+ Aut(T) such that corep(cp-1(Inn(T)) _ {lp}. Define
B := {f:P-+ T: f(pq)=
all peP, qEQ}.
Then B is a group under pointwise multiplication and B = Tk. Let P act on B by f '(x) := f (px) for all p, x E P.
We define T twr,p P to be the semidirect product of B by P relative to this conjugation action of P. Such a twisted wreath product T twr,, P has a transitive action on B such that B acts by right multiplication and for f E B and
pEP,p:fHfP.
TW Q=T k = B, where k > 1 and T is a finite nonabelian simple group, G is a twisted wreath product T twr, P defined as above, and G acts on S2 with the action defined above.
The differences between the conditions on P and Q for primitivity and quasiprimitivity of T twr,o P are rather subtle. A discussion can be found in [26, Remark 2.1], and in [1, Section 5]. We note in particular that for the quasiprimitive type TW we do not require that the image of cp contains Inn(T). For the final quasiprimitive type PA, G preserves a product structure Ak on a G-invariant partition of 0 and the subgroup of Sym(A) involved is quasiprimitive of type AS with socle T. Thus a quasiprimitive group G of type PA
induces a faithful product action on this partition of ft For a group R and positive integer k > 2, a subdirect product of Rk is a subgroup H of Rk such that, for each i = 1, . . . , k, the projection map 7r2: Rk -* R onto the i1h direct factor is surjective when restricted to H, that is, (H)ir = R.
PA N = Tk < G < H wr Sk < Sym(i) wr Sk, where k > 1 and T is a finite nonabelian simple group, H is a quasiprimitive permutation group
on A of type AS with non-regular socle T, and G acts transitively by conjugation on the simple direct factors of N. Choose J E A and set
R := T5. Then 1 < R < T. There is a G-invariant partition 1' of 1 (which is possibly trivial in that the parts of Sl' may have size 1) such that for some w E 1, NN, = Rk and for a E w, Na is a subdirect product
of R. k
Cheryl E. Praeger
76
Of all the quasiprimitive types, this last type PA differs the most from its corresponding primitive type. For a primitive group G of type PA, the partition SZ' is trivial and so may be identified with SZ, and the group H is primitive of type AS. It is clear from the descriptions of the eight types that a finite quasiprimitive permutation group belongs to at most one of these types. The main theorem in [26] shows that the converse is also true.
Theorem 5.1 ([26, Theorem 1]) Each finite quasiprimitive permutation group is permutationally equivalent to a quasiprimitive group in exactly one of the quasiprimitive types HA, HS, HC, AS, SD, CD, TW, PA. 6
Finite quasiprimitive 2-arc transitive graphs
In order to test the effectiveness or power of our observations about the role of quasiprimitive graphs in the families F'2-arc, .F1-prim and Tl-gprim, we might
try to apply the structure result, Theorem 5.1, to study the possible structures of the quasiprimitive graphs in these families. Here we report on the results of following this strategy for the family .F2-a,., of finite vertex-transitive, 2-arc transitive graphs. The first analysis of quasiprimitive graphs in this family using Theorem 5.1 showed that only half of the quasiprimitive types could occur as quasiprimitive 2-arc transitive subgroups of automorphisms.
Theorem 6.1 ([26, Lemmas 5.2 and 5.3]) If r = (V, E) is a finite (G, 2)arc transitive graph such that G is quasiprimitive on V, then G has quasiprim-
itive type HA, AS, TW, or PA.
Further, it was observed in [26, Section 6] that, for each of the types HA, AS, TW, or PA, there are examples of (G, 2)-arc transitive graphs with G quasiprimitive of the given type. The next step in the study of quasiprimitive 2-arc transitive graphs was to investigate further the nature of the graphs corresponding to each of these quasiprimitive types. It was in fact possible to complete the classification of those of type HA. This was done in [13] in joint work with Sasha Ivanov. Recall that all quasiprimitive permutation groups of type HA are primitive.
Theorem 6.2 ([13]) If r = (V, E) is a finite (G, 2)-arc transitive graph of valency n such that G is (quasi)primitive on V of type HA, then IVY = 2d, G = Z .G,, < Zd.(Aut(F) n GL(d, 2)) with Go irreducible, and r, d, n, and 2 2 Aut(F) n GL(d, 2) are as in one of the lines of Table 1.
-
If r = (V, E) is a finite primitive (G, 2)-arc transitive graph of type HA and of valency n, then it was shown in [13] that G must be of the form G = N.G0,
where the normal subgroup N = Z acts regularly on vertices, and the subgroup G,, (which is the stabilizer of the identity element of N) has a faithful 2-transitive action of degree n. Furthermore it was shown that N may be
Finite quasiprimitive graphs
77
F
d
n
Aut(F) n GL(d, 2)
Kn.+1
log2(n + 1)
n
GL(d, 2)
-
i Q. 2
n-1
n odd
Sn
-
am
Pm(a)
m°
U(q)
2 am
-1
Comments
m>3
2a-1
2a)
> q2 - q + 1
q3 + 1
PFU(3, q)
q
r'(C23)
11
23
M23
-
F(C22)
10
22
M22.2
-
3 (mod 4); see below
Table 1: Quasiprimitive 2-arc transitive graphs of type HA.
identified with a quotient of the natural Z2Go permutation module Z for this 2-transitive action of G. by some maximal submodule W such that Ga is faithful on 7G2/W, and moreover that N and hence W determine F. In fact we have V = N = 7G2 /W and {x + W, y + W1 is an edge if and only if x + y - e (mod W) for some weight 1 vector e E Z . Thus, for a 2-transitive permutation group G,, of degree n, the finite primitive (G, 2)-arc transitive graphs of type HA, with point stabilizer Ga, are in one-to-one correspondence with the maximal Z2G,, submodules W of the 7G2Ga permutation module Z
such that Ga is faithful on 7L2/W, or alternatively on the minimal faithful Z2Go submodules of Z n. Most of the examples in Theorem 6.2 are therefore defined in terms of a maximal submodule of the Z2G,, permutation module for the 2-transitive group Ga. Occasionally the examples turn out to be isomorphic to some well-known graphs, and in such cases we use their more familiar descriptions. This is the case for the complete graphs in line 1 of the table, and is also the case for the graphs in line 2. The graph a Qn is the folded cube, the antipodal quotient of the n-dimensional cube; it may be defined as the graph with vertex set V = 7G2-1 with {x, y} an edge if and only if x - y has weight
either 1 or n - 1. The graph Pm(a) (m > 3, a > 1) is defined in terms of the 7G2Go permutation
module Z for Ga where PSL(m, 2a) < Ga < PSL(m, 2a) with Ga acting on the n = (2a - 1)/(2a - 1) points of the projective geometry PGm_1(a). Let W be the Z2Go submodule generated by the characteristic functions of all the complements of hyperplanes of PGm_1(a). Then the vertex set of Pm(a) is the quotient module V := 7G2/W1 with edges as described above.
78
Cheryl E. Praeger
The examples in line 4 probably do not exist (as quasiprimitive examples). When q - 3 (mod 4), the submodule structure of the 7G2G, permutation mod-
ule V = Z for the 2-transitive unitary groups Go, where PSU(3, q) < G,, < PrU(3, q) and n = q3 + 1, has not been completely determined. It is believed, but has not yet been proved, that V has no faithful minimal Go-submodule and therefore provides no quasiprimitive examples in Theorem 6.2. This has been proved by Jane McCorkindale [21] in her D. Phil. thesis in the case where q - 3 (mod 8) and q + 1 # 0 (mod 3), and has also been verified computationally when q = 3, 7, and 11 by Andreas Brouwer (private communication).
The graphs r(C23), r(C22) are constructed in a similar manner from the Z2-permutation modules for the 2-transitive groups M23 and M22.2 respectively. Descriptions of these graphs, including distance diagrams, may be found in [5, Theorems 11.3.4 and 11.3.5], see also [13, Description 1.8]. We next turn to the classification of (G, 2)-arc transitive graphs where G is quasiprimitive of type AS, say T < G < Aut(T) for a finite nonabelian simple group T. On the one hand, for some classes of Lie type simple groups T of low Lie rank it should be possible to classify all (G, 2)-arc transitive graphs. This has been achieved for the classes PSL(2, q), Sz(q) and Ree(q) in [7, 8, 9, 11]. Each of these classifications involved the constructions of several new infinite families of examples, and gave new insights into the nature of such
graphs and in particular of their primitive quotients. On the other hand it can be demonstrated that such a complete and explicit classification is not possible for the family of finite alternating groups T = An. A study was made in [29] of primitive permutation representations of G = An or Sn with a 2-transitive suborbit. All examples were classified except those for which a point stabilizer G,, is a primitive permutation group of degree n of primitive type AS and has a faithful 2-transitive permutation representation. Although all possibilities for the 2-transitive group Gc, are known (using the classification of the finite simple groups, see [6]) we still do not know all faithful primitive representations of all such groups. Thus it is not feasible at present to complete the classification in [29], and consequently it is not feasible to classify even the (G, 2)-arc transitive graphs with G = An or Sn and G primitive on vertices. Robert Baddeley [1] has made a study of finite (G, 2)-arc transitive graphs with G quasiprimitive or primitive on vertices of type TW. There are many examples and his paper gives a general approach to their construction. The situation for finite (G, 2)-arc transitive graphs with G quasiprimitive of type PA is not quite so clear. Examples may be constructed as follows. However we do not know what "typical" quasiprimitive graphs of this type are like.
Example 6.3 Let A = {1,.. . , n} and let H = Sym(O). Then the transposition h = (12) is such that H1,2 = CH(h) fl H1. Let L := (H x H).(7r), where 1r2 = 1 and (h1i h2)r = (h2, h1) for all hi E H. Then L =' H wr Z2. Also let K = {(x, x) : x E H1} and g = (1H, h)lr. Define I' to be the graph with vertex set V = {Ky : y E L} such that {Ky, Ku} is an edge if and
Finite quasiprimitive graphs
79
only if yu-1 E KgK. Now K n K9 = {(x, x) : x E CH(h) n Hl = H1,2} and G :_ (K, g) has index 2 in L. It follows from [8, Theorem 2.1] that the connected component I" of r containing the vertex K is a connected (G, 2)-arc transitive graph and G is quasiprimitive of type PA on the vertices of P. 7
Full automorphism groups of quasiprimitive graphs It is immediate from the definition of primitivity that if G is a primitive
permutation group on S2 and G < H < Sym(SZ), then H also is primitive on Q.
However this implication is not true if G is an imprimitive quasiprimitive permutation group on Q. Indeed if P is a nontrivial G-invariant partition of Q and if H is the subgroup of all permutations of Il preserving P, then G < H < Sym(Sl), H is a wreath product Sym(p) wr Sym(P) (where p E P) and the base group Sym(p)IPI of H is intransitive on Q. Thus an overgroup in Sym(SZ) of a quasiprimitive group may not be quasiprimitive. Recall that, for a family .F of arc-transitive graphs with a group theoretic defining property Q,
a quasiprimitive graph in F is a graph F for which there exists G < Aut(F) such that G is quasiprimitive on vertices and also G has property Q. It may be the case that r is G-quasiprimitive, but Aut(F) is not quasiprimitive on vertices.
For most of the properties we considered in Section 4, if G < H < Aut(F) and G has property Q, then H also has property Q, that is the property Q is inherited by overgroups in Aut(F). (This is true for Fl-a,,, F2-,,r,, and -Fi-prim, but not for .Fi_gprim.) These observations suggest a number of interrelated questions about overgroups in Aut(F) of a given quasiprimitive subgroup G. We pose them in terms of the full automorphism group Aut(F), but they are
equally interesting with Aut(F) replaced by an arbitrary overgroup of G in Aut(F). Questions 7.1 Suppose that FQ is the family of finite arc-transitive graphs r with a certain group theoretic defining property Q which is inherited by overgroups in Aut(F). Let r = (V, E) be a G-quasiprimitive graph in FQ, that is, G < Aut(F), G is quasiprimitive on V, and G has property Q.
(a) Under what conditions can we be certain that Aut(F) is quasiprimitive on V? (b) If Aut(F) is quasiprimitive on V, is it possible that G and Aut(F) have different quasiprimitive types, and if so what are the possible pairs of quasiprimitive types for G, Aut(F)? (c) If G and Aut(F) are quasiprimitive with the same quasiprimitive type, is it possible that soc(G) # soc(Aut(F)), and if so what are the possibilities for these socles?
80
Cheryl E. Praeger
We suspect that, in many of the cases in which Aut(r) is not quasiprimitive on V, the centralizer of G, or of soc(G), in Aut(r) may be a nontrivial intransitive normal subgroup. To be more precise suppose that G is maximal by inclusion in Aut(r) such that G is quasiprimitive on V, and that G # Aut(r).
Let G < H < Aut(r) with G a maximal subgroup of H. Then H has an intransitive minimal normal subgroup M such that M fl G = 1 and H = GM.
(d) Under what conditions is it true that H = G x M? If this is the case, then the approach described in this paper to analysing the structure of r would suggest that we study the normal quotient rM since often r is a multicover, or even a cover, of FM, and rM is (G/M)-quasiprimitive. On the other hand if G acts nontrivially on M (by conjugation), it is often possible to prove that M is elementary abelian and can be identified with a faithful irreducible G-module.
(e) What can we say about r, G and M if M is a faithful irreducible G-module? Many of these questions arose in the study of finite 2-arc transitive graphs. As we have already discussed, the study of finite 2-arc transitive graphs highlighted the importance of finite quasiprimitive permutation groups and motivated the first study of their structure in [26]. They also prompted the questions above. Again we were in a situation where there was insufficient knowledge available about the structure of quasiprimitive permutation groups on a finite set Il and their overgroups in Sym(Sl), even to begin to answer Questions 7.1 (a)-(c). The information was however available for finite primitive permutation groups. The results of [16] give a classification of the overgroups in Sym(Sl) of primitive groups of type AS, while the paper [25] completes the classification of all inclusions G < H < Sym(1) with G primitive on SZ and either G, H of different primitive types, or G, H having different socles. A study applying these results to the automorphism groups of vertex-primitive graphs is in progress [20]. It describes the pairs of groups G, H such that
G < H < Aut(r), G is vertex-primitive, and G, H have different primitive types or different socles. A start has been made in [3] to classifying permutation groups G, H such
that G < H < Sym(Sl), G is quasiprimitive and imprimitive on 1, and H is primitive on Q. A summary statement of the results of [3, 16, 25] may be made as follows.
Theorem 7.2 ([3, 16, 25]) Suppose that Sl is a finite set and that G < H < Sym(Q) with G quasiprimitive on fl and H primitive on S2, such that either G is imprimitive, or G, H are primitive of different primitive types, or G, H are primitive of the same type but with different socles. Then either (a) the pair (G, H) belongs to an explicit list of families of examples; or
Finite quasiprimitive graphs
81
(b) H is of type AS and G is imprimitive; or (c) H is of type PA and G is imprimitive of quasiprimitive type TW or PA. The results in [3] required the solution of several classification problems of special kinds of factorisations of almost simple groups (completed in [2]) and a special investigation (in [4]) of the structure of quasiprimitive groups obtained from the blow-up construction introduced by Laci Kovacs [14]. To extend the classification in Theorem 7.2 it is necessary to solve several more classification problems for almost simple groups (see [3, Section 6]). To
explain this, suppose that G < H < Sym(SZ) with G quasiprimitive and imprimitive, and H primitive. Since all primitive inclusions K < H, with K, H of different primitive types, or of the same type but with different socles, have been classified in [16, 25], we may assume that each subgroup K such that G < K < H is imprimitive. Then by Theorem 7.2 we may assume that H has primitive type AS or PA. Much is known about the case where H has type PA from [4], but some delicate questions remain. From now on let us suppose that H has type AS, with socle T. We are interested only in the case where soc(G) T, as otherwise G would be quasiprimitive of the same type, and with the same socle as H. Let K be a subgroup of H containing G which is maximal such that K 2 T. Then, for a E SZ, since G is transitive on S2, H = GHa, and hence H = KHa is a (so-called max+ or max-) factorisation of the almost simple group H. All possibilities for such triples (H, K, Ha) are classified explicitly in [18, 19]. This classification is not of course sufficient to solve our problem as we need to find those triples for which the subgroup K contains a quasiprimitive subgroup G, and to classify all possibilities for the quasiprimitive types of such subgroups G. As a first step let us suppose that G is maximal by inclusion among the quasiprimitive subgroups of K. If K itself is quasiprimitive then of course we will have (H, G, Ha) in the subfamily of triples (H, L, Ha) with L < H and L quasiprimitive. Suppose now that K is not quasiprimitive and let M be a minimal normal subgroup of K which is intransitive on Q. Since M is intransitive and G is quasiprimitive, we have both H # MHa and M fl G = 1. In particular IH : H, ,I = I Q I divides I G I, which divides I K : MI. Thus the following are some of the preliminary problems whose solution will help to extend the classification in Theorem 7.2.
Problem 7.3 Let H be an almost simple group with socle T. Classify all factorisations H = AB such that B is a maximal subgroup of H and B 2 T, and (a) A = Sd for some nonabelian simple group S and positive integer d, and
AnB=1;or
(b) A is maximal in H subject to A 2 T, A has a unique minimal normal subgroup M and H = MB; or
Cheryl E. Praeger
82
(c) A is maximal in H subject to A
T, A has a minimal normal subgroup M such that H :A MB and IH : BI divides IA: MI.
Recall that a quasiprimitive group has either a unique minimal normal subgroup, or two regular nonabelian minimal normal subgroups. Thus we could relatively easily find all the triples (H, L, Ha), with L < H and L quasiprimitive, from solutions to Problem 7.3 (a) and (b). Problem 7.3 (c) should yield sufficient information for us to locate most of the triples (H, L, Ha), with L < H and L not quasiprimitive but such that L has a quasiprimitive subgroup G. Of course even these classifications would not lead to a complete solution of the problem of determining all inclusions of imprimitive quasiprimitive permutation groups in primitive groups, but they seem a reasonable strategy to pursue. There remains of course the problem of classifying the imprimitive quasiprimitive inclusions.
Problem 7.4 For a finite set S2, classify all inclusions G < H < Sym(S2) such that G, H are both quasiprimitive on Q and of different quasiprimitive types.
Let us return to Questions 7.1 on automorphism groups of quasiprimitive graphs. The group theoretic analysis just described should be applied to give some answers to Questions 7.1 (a)-(c). With regard to Question 7.1 (d), it is certainly possible for a (G, 2)-arc transitive graph F, with G quasiprimitive, to have Aut(F) = G x C with C some small nontrivial cyclic subgroup. An example of such a graph with G quasiprimitive of type TW was found by Robert Baddeley [1, Section 6], and he commented there that such examples seem difficult to find. Recently Cai Heng Li [15] has found an infinite family of examples with G quasiprimitive of type AS. On the other hand, if r = (V, E) is G-locally primitive with G quasiprimitive of type AS, with T := soc(G)
not regular on V and CA t(r) (T) = 1, then it has been shown in [10] that either Aut(I') < AutT, or G < Y < Aut(F) with Y almost simple having soc(Y) # T, or T is a member of a restricted family of Lie type simple groups over a field of characteristic p and Aut(F) contains a semidirect product N.G with N = 7Gd a specific faithful absolutely irreducible 7LpG-module. These observations and results about automorphism groups of graphs r are
simply sample, or preliminary answers to the questions posed above. Much more work remains to be done.
References
[1] Robert W. Baddeley, Two-arc transitive graphs and twisted wreath products, Journal of Algebraic Combinatorics, 2 (1993), 215-237.
[2] Robert W. Baddeley & Cheryl E. Praeger, On classifying all full factorisations and multiple-factorisations of the finite almost simple groups, Technical Report No. 1996/13, University of Leicester, 1995.
Finite quasiprimitive graphs
83
[3] Robert W. Baddeley & Cheryl E. Praeger, On primitive overgroups of quasiprimitive permutation groups, Technical Report No. 1996/14, University of Leicester, 1996.
[4] Robert W. Baddeley & Cheryl E. Praeger, Expansion and blow-up decompositions of quasiprimitive permutation groups, in preparation, 1996. [5]
A. E. Brouwer, A. M. Cohen & A. Neumaier, Distance-Regular Graphs, Springer-Verlag, Berlin, Heidelberg (1989).
[6] Peter J. Cameron, Finite permutation groups and finite simple groups, Bulletin of the London Mathematical Society, 13 (1981), 1-22. [7] Xin Gui Fang, Construction and classification of some families of almost simple 2-arc transitive graphs, Ph. D. Thesis, University of Western Australia, 1995.
[8] Xin Gui Fang & Cheryl E. Praeger, Finite two-arc transitive graphs admitting a Suzuki simple group, Research Report, University of Western Australia, 1996. [9]
Xin Gui Fang & Cheryl E. Praeger, Finite two-arc transitive graphs admitting a Ree simple group, Research Report, University of Western Australia, 1996.
[10] Xin Gui Fang & Cheryl E. Praeger, On graphs admitting arc-transitive actions of almost simple groups, Research Report, University of Western Australia, 1996.
[11] Akbar Hassani, Luz Nochefranca & Cheryl E. Praeger, Finite two-arc transitive graphs admitting a two-dimensional projective linear group, Research Report, University of Western Australia, 1995.
[12] A. A. Ivanov, Distance-transitive graphs and their classification, in Investigations in the Algebraic Theory of Combinatorial Objects (eds. I. A. Faradzev et al.), Kluwer, Dordrecht (1994), pp. 283-378. [13] A. A. Ivanov & Cheryl E. Praeger, On finite affine 2-arc transitive graphs, European Journal of Combinatorics, 14 (1993), 421-444.
[14] L. G. Kovacs, Primitive subgroups of wreath products in product action, Proceedings of the London Mathematical Society, 58 (1989), 306-322. [15] Cai Heng Li, A family of 2-arc transitive graphs, in preparation. [16] M. W. Liebeck, Cheryl E. Praeger & Jan Saxl, A classification of the maximal subgroups of the finite alternating and symmetric groups, Journal of Algebra, 111 (1987), 365-383.
84
Cheryl E. Praeger
[17] M. W. Liebeck, Cheryl E. Praeger & Jan Saxl, On the O'Nan-Scott Theo-
rem for finite primitive permutation groups, Journal of the Australian Mathematical Society, Series A, 44 (1988), 389-396.
[18] M. W. Liebeck, Cheryl E. Praeger & Jan Saxl, The Maximal Factorisations of the Finite Simple Groups and their Automorphism Groups, Memoirs of the American Mathematical Society, 86, No. 432, American Mathematical Society, Providence, Rhode Island (1990). [19] M. W. Liebeck, Cheryl E. Praeger & Jan Saxl, On factorisations of almost simple groups, Journal of Algebra, in press.
[20] M. W. Liebeck, Cheryl E. Praeger & Jan Saxl, On the automorphism groups of finite vertex-primitive, edge-transitive graphs and directed graphs, in preparation.
[21] Jane McCorkindale, The 2-modular representation theory of PSU3(q), q - 3 (mod 4), D. Phil. Thesis, University of Oxford, 1990. [22] B. H. Neumann, Twisted wreath products of groups, Archiv der Mathematik, 14 (1963), 1-6. [23] Peter M. Neumann, Gabrielle A. Stoy & Edward C. Thompson, Groups and Geometry, Oxford University Press, Oxford (1994).
[24] Cheryl E. Praeger, Imprimitive symmetric graphs, Ars Combinatoria, 19A (1985), 149-163. [25] Cheryl E. Praeger, The inclusion problem for finite primitive permutation groups, Proceedings of the London Mathematical Society, Series 3, 60 (1990), 68-88. [26] Cheryl E. Praeger, An O'Nan-Scott Theorem for finite quasiprimitive permutation groups and an application to 2-arc transitive graphs, Journal of the London Mathematical Society, Series 2, 47 (1993), 227-239.
[27] Cheryl E. Praeger, On a reduction theorem for finite, bipartite, 2-arc transitive graphs, Australasian Journal of Combinatorics, 7 (1993), 2136.
[28] Cheryl E. Praeger, Jan Saxl & Kasuhiro Yokoyama, Distance transitive graphs and finite simple groups, Proceedings of the London Mathematical Society, Series 3, 55 (1987), 1-21.
[29] Cheryl E. Praeger & Jie Wang, On primitive representations of finite alternating and symmetric groups with a 2-transitive subconstituent, Journal of Algebra, 180 (1996), 808-833.
Finite quasiprimitive graphs
85
[30] L. L. Scott, Representations in characteristic p, in Santa Cruz Conference on Finite Groups, Proceedings of Symposia in Pure Mathematics, 37, American Mathematical Society, Providence, Rhode Island (1980), pp. 318-331.
[31] D. H. Smith, Primitive and imprimitive graphs, Quarterly Journal of Mathematics, Oxford, Series 2, 22 (1971), 551-557. [32] M. Suzuki, Group Theory I, Springer-Verlag, Berlin (1982).
Department of Mathematics University of Western Australia Nedlands WA 6907 Australia
[email protected] Tree Width and Tangles: A New Connectivity Measure And Some Applications B. A. Reed Summary We discuss tree width, a new connectivity invariant of graphs defined by Robertson and Seymour. We present a duality result and a canonical decomposition theorem tied to this invariant. We also discuss a number of applications of these results, including Robertson and Seymour's Graph Minors Project. 1
1.1
Introduction A taste of things to come A graph is a set of vertices and an adjacency relation which indicates which
pairs of vertices are joined by an edge. Thus, graph theory is essentially the study of connectivity. How then does one measure the connectivity of a graph? Measuring the connectivity between two vertices is straightforward. Two
vertices are said to be k-connected if there are k internally vertex disjoint paths between them. A classical theorem of Menger [30] states that vertices a and b are k-connected in a graph G precisely if there is no set X of fewer than k vertices such that a and b lie in different components of G - X. Standard alternating paths techniques, see e.g. [21], allow us to find either k internally vertex disjoint a-b paths or such a set X efficiently. An appropriate definition of a highly connected graph, or of a highly connected piece of a graph is more difficult. The classical approach is to call a graph k-connected if every pair of its vertices is k-connected. This definition, although natural, does not capture the kind of connectivity that will concern us. It focuses on local properties rather than global ones. To illustrate what we mean, we consider the hexagonal lattice C (see Figure 1). G has maximum degree three and hence contains no 4-connected subgraph. On the other hand, if X is a set of at most three vertices of G then G - X is either connected or has two components, one of which is a vertex. Similarly, if X is a set of at most k vertices then the largest component of G - X contains all but at most k2 of the vertices of G. Thus, no cutset of size k "globally disconnects" L. The same is true for sufficiently large and robust finite subgraphs of G. One of the goals of this paper is to introduce its readers to a connectivity invariant, the bramble number, which measures global connectivity. Having done so, we will examine those graphs for which this invariant is bounded by a given k. If the bramble number of G is at most two then G is a forest. It turns out that for any k, the graphs with bramble number at most k
have a "tree-like" structure. How tree-like depends on the value of k. We will present a precise duality theorem after we have developed the requisite definitions. Now, such tree-like structures are easy to deal with. For example,
B. A. Reed
88
Figure 1: The hexagonal lattice we will develop efficient algorithms for resolving seemingly difficult (i.e. NPcomplete) problems on graphs whose bramble number is bounded. We will
also see that many difficult conjectures can be resolved when restricted to such graphs. We will also examine the structure of graphs with large bramble number.
It turns out that if a graph has large bramble number then it must contain a subdivision of a large piece of the hexagonal lattice. This fact allows us to resolve many difficult questions on such graphs. Combining the two complementary types of results discussed above has proven very fruitful. We will present some of the many results obtained using this technique. We will also discuss an assortment of other results which elu-
cidate the dual notions of brambles and tree-structures. The most important of these is a theorem which states that every graph can be decomposed into a tree structure of pieces each of which corresponds to a maximally connected part of the graph. We flesh out this skeletal overview in the next four subsections, developing
the necessary definitions en route. The theory of brambles and tree decompositions (to use their proper name) was developed mainly by Robertson and Seymour ([36, 37, 38, 41]). The first four sections of this paper survey the development of the theory. We discuss applications of the theory in Sections 5 and 6. Section 5 presents the applications that Robertson and Seymour had in
Brambles And Tree Decompositions
89
mind when they developed the theory. Section 6 discusses some more recent applications of the theory to packing and covering problems. These results demonstrate that brambles and tree decompositions are not just specialized pieces of mathematical machinery but rather a new central theory which is ripe for further development. In what follows, we may confound a graph with its vertex or edge set for notational convenience. We will often use X + y to denote X U {y}. Also, we will often use n for IV(G) I particularily when measuring computational complexity. 1.2
Measuring connectivity
In this section, we present a number of invariants which measure the global
connectivity of a graph. We shall take as our archetypes of globally connected graphs, subdivisions of certain finite subgraphs of the hexagonal lattice. Specifically, we shall consider walls.
Figure 2: An elementary wall of height 8
An elementary wall of height 8 is depicted in Figure 2. An elementary wall of height h is similar. It is a piece of the hexagonal lattice consisting of h levels each containing h bricks. More precisely, an elementary wall of height h contains h + 1 vertex disjoint paths, R1, ... , Rh+1 which we call rows, and h + 1 vertex disjoint paths, C1, ..., Ch+1 which we call columns. The reader should be able to complete the definition by considering Figure 2, in which R1 is the top row. (For fussy formalists: the first and last row, i.e. R1 and Rh+1, both contain 2h + 1 vertices. All the other rows contain 2h + 2 vertices. All the columns contain 2h vertices. Column i joins the (2i - 1)st vertex of R1 with the (2i - 1)st vertex of Rh+1; it contains the (2i - 1)st and 2ith vertex of every other row, as well as the edge between them. For j < h
B. A. Reed
90
and odd, each Cz contains an edge between the (2i - 1)st vertex of Rj and
the (2i - 1)st vertex of Rj+1. For j < h and even, each Ci contains an edge between the 2ith vertex of Rj and the 2ith vertex of Rj+l. These are all the
edges of the wall.)
The perimeter of an elementary wall of height h is the cycle formed by R1 U Cl U Rh+1 U Ch+l. Its corners are: R1 n C1, R1 n Ch+l, Rh+1 n Cl, and Rh+l n Ch+1 A wall of height h is obtained from the elementary wall by replacing the edge
set by a corresponding set of internally vertex disjoint paths whose interiors are vertex disjoint from the original elementary wall, see Figure 3. The rows, columns, corners, and perimeter of the wall correspond to the same objects in the original elementary wall. The nails of the wall are the vertices of degree three within it as well as its corners.
I
40--*
1 Figure 3: A wall of height 3
We say that a set S of vertices in a graph G is k-linked (k an integer) if for any set X of fewer than k vertices there is a (unique) component of G - X containing more than half of the vertices of S. Obviously, the vertices of any clique C are [IV(C)I/2] -linked. We note that the nails of any wall of height h can be shown to be [h/2]-linked (since any set of at most h/2 vertices misses half the rows and the nails in these rows will clearly all be in the same component), even though this set may be independent. One measure of the connectivity of G which interests us is the largest k for which G contains a k-linked set. We shall denote this by linkedness(G), the linkedness of G. We now introduce a measure of connectedness which is more abstract. To motivate it, let us consider a k-linked set S in a graph G. For each subset X of V (G) with IX I < k, the big component of G - X is that which contains more than half the vertices of S. Let ,3 be the set of big components. Obviously any two elements of 0 intersect. Just as obviously, no set X of fewer than k vertices intersects every element of 0. By a bramble, we mean a set of connected subgraphs any two of which touch, that is intersect or are joined by an edge. We note that the vertices of
Brambles And Tree Decompositions
91
any clique form a bramble. Since every row of a wall intersects every column, a wall of height h contains the bramble Crossesh = {R U C I R is a row, C is a column}.
The order of a bramble a is the minimum cardinality of a hitting set for 8 (i.e. a set X of vertices which intersects every element of 3). Obviously the order of the bramble given by the vertices of a clique is just the cardinality of the clique. The order of the bramble Crossesh is h + 1. To see that it is at most h + 1, we note that taking one vertex from each row yields a hitting set. To see that it is at least h + 1, we note that if X is a set of less than h + 1 vertices then X misses some row R and some column C and hence the corresponding element of Crossesh. The bramble number of G, denoted BN(G) is the maximum of the orders of its brambles. Our remarks in the last paragraph show that every k-linked set generates a bramble of order k and hence for every graph G: BN(G) > linkedness(G). Before going any further, we prove that a partial converse holds.
Theorem 1.1 For every graph G, linkedness(G) < BN(G) < 2 linkedness(G).
Proof We know BN(G) > linkedness(G). Therefore we need only show that linkedness(G) > BN(G)/2. Clearly, it suffices to show that if X is a minimum order hitting set for a bramble /Q then it is [IXI/21-linked. Suppose not, and let Y be a set of fewer than IXI/2 vertices such that no component of G - Y contains more than half the vertices of X. Since IYI < IXI, Y is not a hitting set for /Q, so there is a component U of G - Y which contains an element of /3. Since every two elements of Q touch, it follows that every element of /Q intersects U U Y (and hence U is unique). Since X is a hitting set for ,13, X fl U is a hitting set for those elements of /3 completely contained in U. Thus (X fl U) U Y is a hitting set for (3, contradicting the minimality of IXI. Now, Theorem 1.1 shows that our two connectivity parameters are essentially measuring the same property of graphs. A much more difficult theorem shows that the property they measure is tied to the size of the largest wall in the graph. To wit:
Theorem 1.2 ([46] (see also [38])) Let h be the maximum of the heights of the walls in G. Then h + 1 < BN(G) < 2534"5
We have already proven that h + 1 < BN(G). We need only prove that the second inequality holds. However, before proving this theorem characterizing
graphs whose bramble number is high, we analyze the structure of graphs whose bramble number is low. We sketch a proof of Theorem 1.2 in Section 4 and give the details in Section 7.
B. A. Reed
92
1.3
Tree decompositions and tree width
A tree decomposition of a graph G consists of a tree T and for each node t of T, a subset Wt of V (G) such that:
(i) For each vertex v of G, the set S = It t is a node of T, v is in Wt} I
induces a non-empty subtree of T, and (ii) For each edge of G with endpoints x and y, S,, intersects Sy.
We let W = (Wt
I
t is a node of T), and speak of the tree decomposition S if this precision is necessary.
[T, W]. We use
Wt, = {a, b, d}
Wt2 = {c, b, e}
W,3 = Id, e, f} Wt, = {b,d,e}
W
G
Xt, =
:\b
Xt2 =
b
e
X t3
T and the ST,
X
Figure 4: An example of a tree decomposition
As depicted in Figures 4 and 5, given a tree decomposition [T, W] of G, we can choose for each node t in T, a subgraph Xt of G with node set Wt such that each edge of G is in precisely one of these subgraphs. To do so, we place each edge e with endpoints x and y in Xt for some arbitrary element t of S,, n Sy. For any such set of subgraphs, we set X = (Xt I t is a node of t) and speak of the tree decomposition [T, X]. We justify this abuse of notation by remarking that we are really always considering a partition of the edge set and hence a tree decomposition of the second type. However, sometimes we simply
Brambles And Tree Decompositions
V1
93
Vk_1 Vk
V2
Wt; _ {vi-k, vi-k+1, ... , vi}
.
Vk+2
v2k-1
Vk+1
for k < i < k2
V2k
Wti = {v1,v2,...,vi}
for1 BN(G) - 1, we shall prove TW(G) < BN(G) - 1 in Subsection 2.3. Actually, we now prove the following statement which implies immediately that TW(G) > BN(G) - 1: Fact 1.4 If ,Q is a bramble of G and [T, W] is a tree decomposition of G then there is a node t of T such that Wt is a hitting set for ,Q. Proof We shall need the following two facts:
1.5 (The Helly Property For Trees) If S is a family of at least two subtrees of a tree T, every pair of which intersect, then n{S I S E S} is non-empty.
Brambles And Tree Decompositions
95
Proof We proceed by induction on IV(T)1. If T has only one node then there is nothing to prove. Otherwise, let t be a leaf of T with neighbour s.
If any tree S in S is exactly t then n{S IS E S} is t. Otherwise, T - t is a tree and for each S in S, S - t is a subtree of T - t. Furthermore, every pair of trees in S' = IS - t I S E S} intersect because if t is in S, fl S2 for some Sl, S2 E S then so is s. So by induction the trees of S' have a non-empty common intersection and so do those of S.
Fact 1.6 If C is a connected subgraph of a graph G and [T, W] is a tree decomposition then Sc = It I Wt intersects C} induces a subtree of T.
Proof This is true by definition if C has only one vertex. We proceed by induction on ICI. So, let C be a connected subgraph of G with at least two vertices and let v be a leaf of some spanning tree for C. Then, C - v is connected and by induction, SC_ is a subtree of T. S is also a subtree of T and since there is an edge from v to C - v, we know these two subtrees intersect. Thus the union of their vertex sets, which is precisely SC, induces a subtree of T, as required. Now, let ,(3 be a bramble in a graph G and let [T, W] be a tree decomposition
of G. For each B in ,Q, we let SB be the set {t I Wt intersects B}. We know each SB induces a subtree of T. Now, for every B and C in ,3, B touches C and so SB fl Sc is non-empty. Thus, by the Helly property for trees, there is a node tin n{SB I B E /3}. Now, Wt is a hitting set for 0, as required. In Section 2, in addition to completing the proof of this duality theorem, we will discuss methods for solving difficult optimization problems on graphs of bounded tree width in polynomial, often linear, time. The idea is to generalize standard dynamic programming techniques from trees to bounded-width tree decompositions. These techniques work on trees because removing any edge disconnects the graph. We can generalize the techniques to graphs of bounded tree width because of the following analogous result which shows that the edges of a tree decomposition of G of width k correspond to cutsets of size at most k.
Definition Let [T, W] be a tree decomposition of a graph G. For any subtree S of T, by Vs we mean U{Wt I t is a node of S}. Lemma 1.7 Let [T, W] be a tree decomposition of a graph G. Let rs be an arc of T and let R and S be the components of T - rs containing r and s respectively. Then, (VR - Ws, Vs - W,.) is a partition of V - (W, fl W9) and furthermore, there is no edge of G between VR - W9 and Vs - W,..
Proof First, we note that for each vertex v of G, exactly one of the following
holds: S C R and hence v E VR - W3, S C S and hence v E Vs - W, or S,, contains the are rs and hence v E W3 fl W,.. Thus, (VR - W Vs - W,.) is a partition of V - (W, fl WS). Now, if u is in VR - W,, and v is in Vs - W. then Su C R while S C S. Thus, Su fl S, = 0 and so uv ¢ E(G).
B. A. Reed
96
Corollary 1.8 Let [T, W] be a tree decomposition of a graph G and let t be a . . , si. Let Si be the component of T - t containing si. Then (Vs, -Wt, . . . , Vs, -Wt) is a partition of V -Wt. Furthermore, for 1 < i < j < 1, there is no edge between Vs, - Wt and V53 - Wt. node of T with 1 neighbours s 1 , .
To complete this subsection, we present a number of technical results concerning tree decompositions. Given a tree decomposition [T, W] of a graph G, if there are two adjacent nodes s and t of T such that WS C Wt then we can contract s and t into one new node s * t to form a smaller tree T' (i.e. V (T) = V (T) - s - t + s * t and
E(T') = E(T - s - t) U {x(s * t) I xs or xt E E(T)}). It is easy to see that if we let W,: = W, for each node r of T' except s * t, set Ws*t = Wt , and set W' _ (W; I r is in T') then [T',W'] is also a tree decomposition of G with the same width as [T, W]. So, if we choose a tree decomposition [T, W] of G of minimum width which has no more nodes than any other tree decomposition of G of the same width then for any adjacent nodes s and t of T, we know that W, neither is contained in nor contains Wt. We call a tree decomposition with the latter property nice. We can prove by induction on the number of vertices in G that
1.9 any nice tree decomposition [T, W] of G has no more nodes than G has vertices.
Proof This is clearly true if T has only one node. If T has at least two nodes then we simply find a leaf 1 of T with neighbour t and delete the set Y of
vertices of G which appear in Wt but not in Wt. [T - 1, W - Wt] is a tree decomposition of G - Y, since Wt - Y C_ Wt. It is clearly nice. Furthermore, Y is non-empty since [T, W] is nice. The result follows. We can also use nice tree decompositions to prove
1.10 a simple graph G of tree width at most k has minimum degree at most k. Hence if G has n vertices then it has at most kn edges.
Proof We prove this by induction on the number of vertices in G. It is obviously true if G has a tree decomposition of width at most k using a one node tree. So, we assume that G has tree width at most k and let [T, W] be a nice tree decomposition of G of width at most k with at least two nodes. Again we consider a leaf 1 of T and the necessarily non-empty set Y of vertices which appear in Wt and in no other W3. Clearly, each vertex y in Y has degree
at most k in G as its neighbourhood is contained in W1. The rest of (1.10) follows by induction.
Brambles And Tree Decompositions
97
Tangles, separations, and canonical tree decompositions
1.4
The duality theorem discussed in the last section shows that the notions of brambles and tree-decompositions are intimately linked. We want to investigate this relationship further. To do so we will need to consider some special brambles called tangles. We will also need to look at tree decompositions from a new perspective, for which we will need to introduce the notion of separations. Having developed this machinery, we will be able to state our theorem concerning the existence, for every graph, of a tree decomposition whose nodes correspond to the maximally connected pieces of the graph.
A separation of G consists of an ordered pair (A, B) of subgraphs of G which have disjoint edge sets and whose union is G. The order of a separation (A, B), denoted ord((A, B)), is OV(A) fl V (B) I. A and B are its sides. Two separations (A, B) and (C, D) are laminar if either A C C (and hence
D C B), or A C D (and hence C C B), or B C C (and hence D C A), or B C D (and hence C C A). A set of separations is laminar if every pair of separations within it is. Note that any arc rs in a tree decomposition [T, X] corresponds to two separations (A, B) and (B, A) where, if T1 and T2 are the two components of T - rs, we have:
A=U{XtItisinTl}, B=U{XtItisinT2},
and
V(A fl B) = V(Xr) fl V(XS).
We shall call these two separations, the separations made by rs. It is easy to verify that the set of separations made by the arcs of a tree decomposition is laminar. It is also straightforward to verify that any laminar set S of distinct separations in a graph G corresponds to a unique tree decomposition [TS, Xs] of G such that each separation S in S is made by precisely one arc of TS and
for each arc a of TS one of the separations made by a is in S (this was first noted in [41]). A bramble T is a tangle if for any triple {T1,T2,T3} of elements of T either (i) T1 fl T2 fl T3 is non-empty, or
(ii) there is an edge e such that all of T1, T2, and T3 contain an endpoint of e.
Remark Note that (i) implies (ii) unless Tin T2flT3 is a singleton component. The tangle number of G, denoted TN(G) is the maximum order of a tangle in G (the order of a tangle is its order as a bramble). In [41], Robertson and Seymour define branch width and use a duality theorem relating the tangle number of a graph to its branch width to show that the bramble number of G is at most 3/2 of its tangle number. This bound is best possible, as is shown by the graphs Kk, k > 1 (see [41]). We shall show here:
B. A. Reed
98
1.11 The bramble number of G is at most 3 times its tangle number. To do so, we need some definitions.
A preference of order k in a graph G is a function f mapping each set X of k - 1 or fewer vertices of G to a component of G - X such that if X1 and X2 are both sets of at most k - 1 vertices of G then f (X1) and f (X2) touch. A strong preference of order k in a graph G is a preference f such that
if X1, X2, and X3 are all sets of at most k - 1 vertices of G then either f (X1) fl f (X2) fl f(X3) is non-empty or there is an edge of G which has an endpoint in each of these three sets. Now, if /3 is a bramble of order k, then for each X C V(G) with iX < k there is exactly one component of G - X containing an element of /3. We let f#(X) be this component. Clearly, fp is a preference of order k. Conversely, if f is a preference of order k then
Of = If (X) I X C V(G), IXI < k} is a bramble. Furthermore, /3 f has order at least k because for any set X of fewer than k vertices, X is not a hitting set for of as it does not intersect f (X). Similarly, if T is a tangle of order k then fT is a strong preference of order k while if f is a strong preference of order k then Of is a tangle of order at least k.
Proof of 1.11 Now, consider a bramble /3 of order k in a graph G, and the associated preference fp. We consider the restriction fQ of fp to sets of size at most [k/31 - 1 or less. We claim fp is a strong preference of order [k/3]1 in G and hence that /3 f, is a tangle of order Ik/31 or more. This implies that the bramble number of a graph is indeed at most 3 times as great as its tangle number.
To see that fQ is indeed a strong preference consider three sets X1, X2, X3 in V (G) each with at most Ik/31 - 1 elements. As IX1 U X2 U X31 < k, there is an element B of ,3 which fails to intersect X1 U X2 U X3. But
now, B C_ f#(XI) = f,3(X1). Similarly, B C f'' (X2) and, B C f'(X3). So, f13 (X1) fl f13 (X2) fl fp(X3) is non-empty and fp' is a strong preference as
claimed. So we have proven (1.11).
In fact, we have shown that every bramble of order at least 3k generates a tangle of order k. Thus, we do not lose too much by restricting our attention to tangles. We will therefore do so, because as we are about to see, there is a natural definition of tangles in terms of separations whilst the same is not true of brambles. A bias of order k in a graph G is a function f mapping each separation (A, B) of order less than k either to A or to B such that
(i) V (f ((A, B))) - V (A fl B) 0 (or equivalently: G contains at least k + 1 vertices and if V(B) = V(G) then f ((A, B)) = B) and
Brambles And Tree Decompositions
99
(ii) if Sep, and Sep2 are both separations of order at most k in G then E(f (Sep,)) fl E(f (Sep2)) is non-empty.
A strong bias of order k in a graph G is a bias f such that if Sep,, Sep2, and Sep3 are three separations of order at most k in G then E(f (Sep,)) n E(f (Sep2)) fl E(f (Sep3)) is non-empty.
Y Figure 6: Graph showing that bias is not equivalent to preference Now, a preference g of order > 2 defines a bias f in a natural way: f ((A, B)) will contain g(V(A) fl V(B)) (one of the two sides will contain the big component by the definition of separation). Similarily, a strong preference defines a strong bias. If every bias arose from a preference in this manner then the two notions would be equivalent. Unfortunately, this is not the case. As an example, consider the graph in Figure 6. We define a bias of order two by setting f ((A, B)) to be the side of (A, B) with two or more edges. By considering the separations (A, B) with V (A fl B) = a, we can show that this bias does not arise from a preference in the manner discussed above. What makes tangles so useful is that every strong bias does arise from a strong preference, as we now show.
Lemma 1.12 For every strong bias f of order k there is a strong preference g of order k such that f ((A, B)) contains g(V(A) fl V(B)).
Proof Let f be a strong bias of order k in a graph G. We claim that for every set X of less than k vertices there is a component g(X) of G - X such that letting A be the subgraph consisting of g(X) and the edges of G between g(X) and X, we have: f ((A, G - A)) = A. Using the fact that f is a bias, we see that, for every separation (R, S) of G with V(R) n V(S) = X, f ((R, S)) contains g(X). Since f is a strong bias, g is a strong preference of order k. Thus, f arises from the strong preference g. To prove our claim, we assume the contrary and choose a set X for which the claim is false. We also choose a separation (A, B) with V (A fl B) = X, f ((A, B)) = A, and A minimal subject to these conditions. Assume first that some edge e of A has both endpoints in X. Then, by the minimality of A, f ((A - e, B + e)) = B + e. By the first condition on a bias, f (e, G - e) = G - e. But E(A) fl E(B + e) fl E(G - e) = 0 contradicting the fact that f is a strong bias. So this case cannot occur. Thus, there must be two components Cl and C2 of G - X contained in A.
We let Al = A - V (Cl) and A2 = A - V(C2). By the minimality of A,
B. A. Reed
100
f ((A1, G - A1)) = G - Al and f (A2, G - A2) = G - A2. Once again, we contradict the fact that f is a strong bias, since E(G - A1) n E(G - A2) n E(A) _ 0. This completes the proof of the claim and the lemma.
We say that two brambles X31 and /32 are distinguishable if there is some
X C_ V(G) with JXJ < ord(,Q1), JXi < ord(32) such that fp,(X) # fpz(X). Otherwise, the two brambles are indistinguishable. If 01 and X32 are distinguishable then any set X C_ V(G) with JXJ < ord(i31), JXJ < ord(,32) such that ff, (X) 0 f,32 (X) is a (/01 i ,Q2)-distinguisher. We call a bramble maximal if there is no bramble indistinguishable from it with higher order. We call a tangle maximal if there is no tangle indistinguishable from it with higher order. We have seen that for every tree decomposition [T, X] of a graph G and
every tangle T in G, there is a node t of T such that V (Xt) is a hitting set for T (in fact this holds for all brambles, not just tangles). We can now state the following companion result:
Theorem 1.13 (The Canonical Tree Decomposition Theorem [41]) For any graph G, we can construct a tree decomposition T which has the following properties.
(1) For each maximal tangle T of G, there is exactly one node t(T) of T whose vertices form a hitting set for T.
(2) If Ti and T are indistinguishable maximal tangles (and hence have the same order) then t(Ti) = t(T).
(3) If Ti and T2 are distinguishable maximal tangles then t(Ti) 0 t(T2) and there is an arc st on the unique t(Ti) to t(T2) path of T such that V (XS) n V (Xt) is a (T1, T2)-distinguisher of minimum order.
(4) For every node t of the tree decomposition, there is a maximal tangle T
such that t(T) = t. Intuitively, this tree decomposition splits G up into its highly connected pieces using cutsets which are as small as possible. We note that because each Wt is a hitting set for some tangle for which no
other Ws is a hitting set, there is no s # t such that Wt is contained in W. It follows that this tree decomposition is nice, and hence any set of distinguishable
tangles contains at most n elements. This contrasts with the situation for brambles as a graph may contain exponentially many distinguishable maximal brambles. We delve more deeply into the properties which separate tangles from brambles in Section 3.
Brambles And Tree Decompositions 1.5
101
Graph minors
Readers who have been looking up the references will have noticed that the results discusssed in Subsection 1.3 come mostly from a paper entitled Graph Minors X: Obstructions to tree decompositions, and that a precursor of Theorem 1.2 appears in a paper entitled Graph Minors V. Excluding a planar graph. These are two of a long series of papers written by Robertson and Seymour (presently more than 25 papers in the series have been written or are in preparation) which contains a number of seminal results concerning graph minors. We discuss the role that tree decompositions and brambles played in the development of this theory in Section 5. For the moment, we will content ourselves with making a few definitions and stating the most important results obtained by Robertson and Seymour.
We contract an edge e in a graph G to obtain a new graph Gxy with V (Gxy) = V (G) - x - y + (x * y) and
E(Gxy) = E(G - x - y) U {z(x * y) I zx or zy E E(G)}. H is a minor of G if a graph isomorphic to H can be obtained from a subgraph of G by a sequence of edge contractions. This is equivalent to requiring that for each edge e of H there is some edge im(e) of G and for each vertex v of H
there is some connected subgraph im(v) of G such that for an edge e = uv of H, im(e) has one endpoint in im(u) and the other in im(v). We shall refer to such a structure as a model of H in G. The results which we will discuss are:
1.14 (Wagner's Conjecture) If G1, G2, G3, ... is an infinite sequence of graphs then there exist i
j such that Gi is a minor of G3.
1.15 For any fixed graph H, there is a polynomial-time algorithm which solves the following decision problem:
Problem: H-Minor Containment
Instance: A graph G. Question: Does G contain an H-minor? 1.16 For any fixed positive integer k, there is a polynomial-time algorithm which solves the following decision problem.
Problem: k Disjoint Rooted Paths
Instance:
A graph G and two sets of vertices of G: X = {x1, ... , xk}, Y = {y1, ... , yk}. Question: Are there k vertex disjoint paths P1, ..., Pk in G such that Pi has endpoints xi and yi?
These results have a variety of important corollaries. One of the most interesting is the following:
102
B. A. Reed
1.17 If C is a class of graphs closed under the taking of minors then there is a polynomial-time recognition algorithm for C. Remark Minor Containment is clearly NP-complete if H is part of the input as determining if a graph has a Hamilton Cycle is a special case. Karp [22] proved that Disjoint Rooted Paths is NP-complete if k is part of the input. Lynch [28] showed this remains true even if we restrict our attention to planar graphs. Remark The proof of (1.14) is found in [44]. The proofs of (1.15) and (1.16) are found in [42]. As we shall see, (1.16) implies (1.15).
Remark The algorithms of Robertson and Seymour for (1.15), (1.16), and (1.17) run in O(IV(G)13) time. Reed [34] developed similar algorithms for the same problems which run in o(IV(G)I2) time.
We show now that (1.14) and (1.15) together imply (1.17). Let C be a class of graphs closed under the taking of minors. Then, there is a minimal
set S of graphs such that G E C if and only if for each H E S, G has no H-minor. This set is called the obstruction set for S. (1.14) implies that S is finite. (1.15) implies that for any fixed H E S, we can check for H-minor containment in polynomial time. (1.17) follows. The results we have just stated have hordes of implications, we mention just a few of them here. To do so, we need to define some minor closed classes of graphs. For any surface E, the class Cr of graphs embeddable in E is clearly closed under the taking of minors (in fact, minors were defined by Tutte [51] who used them extensively whilst studying planar graphs). Now, Kuratowski [24] showed that the obstruction set for planar graphs consists of the two graphs K3,3 and K5 depicted in Figure 7. Archdeacon and Huneke [3] proved that for any non-orientable surface E the obstruction set for Cr is finite. Clearly, (1.14) implies that for any surface E, the obstruction set for Cr is finite (this result actually appears in [40]). Furthermore, (1.17) implies that one can test for any fixed surface E whether G is embeddable in E in polynomial time and, by our remarks, in O( I V (G) 12) time. Polynomial time algorithms for this problem had already been developed [18] but the exponent of their running times depended on E. A graph is called linklessly embeddable if it can be embedded in three space so that no pair of disjoint cycles form a link (as do for example two consecutive
links in a chain). Clearly, if a graph is linklessly embeddable so are all its minors. Thus, (1.17) implies that there is an algorithm to determine if G is linklessly embeddable which runs in polynomial time. Previously, we had not even known if any such algorithm existed. More recently, Robertson, Seymour, and Thomas [45] constructed the obstruction set for the linklessly embeddable graphs.
Brambles And Tree Decompositions
K3,3
103
K5
Figure 7: Kuratowski's forbidden minors Other problems to which (1.17) can be applied include: gate matrix layout, topological bandwidth, disk dimension, and vertex integrity, see [16] and [17] for details. We remark that (1.17) only implies the existence of an algorithm to determine if a graph is in a minor closed class C. It does not tell us how to construct such an algorithm, because we do not know how to find the obstructions. If we had some absolute bound on the size of such obstructions, we could find the
obstruction set and construct an algorithm. This fact has generated interest in bounding the size of the graphs in the obstruction set for particular classes of graphs. Tree decompositions play an important role here, see e.g. [47, 25]. In Section 5, we discuss the role tree decompositions and tangles play in the proofs of (1.14) and (1.16). We close this section by sketching a proof that (1.16) implies (1.15). We begin with a remark and some definitions:
Remark An algorithm for k Disjoint Rooted Paths implies an algorithm for the more general problem where we allow the the paths to share endpoints but insist that they are otherwise disjoint (thus X and Y may be multi-sets and may intersect). We simply make multiple copies of any vertex which appears more than once in X U Y.
Definitions To subdivide an edge e in a graph H, we replace it by a path of length two through a new vertex. A subdivision of a graph H, consists of a graph obtained from H by repeatedly sudividing edges. That is a graph in which each edge of H has been replaced by a path with the same endpoints such that these paths can share endpoints but are otherwise disjoint. We say G contains a subdivision of H if there is a subgraph of G isomorphic to a subdivision of H. We refer to such a subgraph as a smodel of H. We refer to the vertices of this smodel which correspond to the vertices of H as the centres of the smodel. We note the following:
104
B. A. Reed
Lemma 1.18 If H is a graph of maximum degree 3 then G has H as a minor if and only if G contains a subdivision of H.
Proof Let H be a graph of maximum degree three. Assume G contains a subdivision of H, and let F be an smodel of H in G. For each edge e of H arbitrarily choose some edge of the path of H corresponding to e to be im(e). Deleting the chosen edges decomposes F into components each containing one centre. For each vertex v of H we let im(v) be the component containing the centre corresponding to v. This yields a model of H in G and hence H is a minor of G. Now, if G contains an H-minor, then we shall consider the contractions performed to obtain H from a subgraph G' of G chosen so that E(G') is minimal. It follows that we never contract an edge with an endpoint of degree one. We consider the contractions in reverse order and as a sequence of "decontractions". We claim that each graph we meet is a subdivision of H. In any decontraction, we take some vertex v of degree at most three and replace it by an edge uw so that u and w together are incident to at most three edges other than uw. It follows that one of them, say w, is incident to only one other edge, say xw. But now, the decontraction simply consists of subdividing the edge vx. Thus, G contains a subdivision of H.
Lemma 1.19 For any graph H, there is a finite set ZH of graphs such that G has H as a minor if and only if G contains a subdivision of some element of ZH.
Proof We again use the idea of decontracting edges. So we consider a minimal subgraph G' of G for which there exists a sequence of decontractions from H
to a graph isomorphic to G'. We see that every time we decontract in F to obtain F either we are simply subdividing an edge or EVEV(F') dF' (v) - 2 < EvEV(F) dF(v) - 2 (where di(v) is the degree of v in J). It follows that we do only EvEV(H) max(0, dH(v) - 2) decontractions which are not subdivisions. Since we can do the subdivisions after all the other decontractions, the result follows.
By (1.19), if we can test if G has a subdivision of any fixed H in polynomial time then we can test if G has any fixed F as a minor in polynomial time. To test if G has H as a subdivision we need only test, for each of the O(nlV(H)I) injections of V(H) into V (G), whether G has an smodel of H where the given injection specifies the centre corresponding to each vertex. To do this, however, we need only solve the extension of k Disjoint Rooted Paths which we remarked
earlier is no more difficult than k Disjoint Rooted Paths. Thus, (1.16) does indeed imply (1.15). We remark that in fact Robertson and Seymour actually proved (1.15) directly using the same technique they use to prove (1.16). They thereby obtain an 0(n3) algorithm for minor containment.
Brambles And Tree Decompositions 2 2.1
105
Graphs of bounded tree width Using tree decompositions of bounded width
In this subsection, we describe how dynamic programming can be used to efficiently solve optimization problems on graphs of bounded tree width. Many
of the algorithms we describe run in linear time. They all require as input a bounded width tree decomposition of the graph. Finding such a decomposition quickly is the subject of the next subsection. To begin, we recall how to use dynamic programming to solve optimization problems on trees.
Figure 8: A rooted tree
Definitions We root a tree T at a node r by replacing each edge e of T by an arc with the same endpoints directed towards the component of T - e not containing r. Thus, for each node t there is now a directed path from r to t. This gives us a rooted tree (T, r) with root r.
Definitions Let s and t be nodes of a rooted tree. If there is an are from s to t then s is the parent of t and t is a child of s. If there is a directed path from s to t containing at least one arc then t is a descendant of s and s is an ancestor of t.
Definition Let s be a node of a rooted tree (T, r). We define T3 to be the rooted tree with root s consisting of s and all its descendants. The following is a well known simple result:
Fact 2.1 There is a linear-time algorithm which given a rooted tree produces an ordering of its nodes in which each node appears before all of its descendants. Reversing this order yields:
B. A. Reed
106
Fact 2.2 There is a linear-time algorithm which given a rooted tree produces an ordering of its nodes in which each node appears after all of its descendants.
It is this last fact which allows us to solve many optimization problems quickly on trees.
To illustrate, we consider finding a maximum weight stable set. More precisely, we assume we are given a tree T, and an integer weight wt for each node t in T. We want to find MWS = max{>,,ES w(v) I S C V(T), S stable}. To do so, we first root T at some arbitrary node r. Then, for each node s of T, we compute two parameters:
w1(s) = max E {w(v) I S C V (T,,), S stable, s E S
and
vES
w2(s) = max Ew(v) j S C V(T3),S stable, s ¢ S vES
We note that MWS = max(w1(r), w2(r)). To compute w1 and w2, we first order the nodes of the tree so that every node appears after its descendants. We will consider the nodes in this order. We note that if s is a leaf, then w1(s) = w(s) and w2(s) = 0. If s has descendants t1, ... , tk then clearly
=
Fk
1 max(wj(ti), W2(ti)). Thus, we can compute w1(s) and w2(s) in 0(k) time. Since the sum of the number of descendants of the nodes in T is simply IV (T) - 1, it follows that MWS can be computed in linear time.
W1 (S) = w(s) + Ei 1 W2 (ti) and w2(s)
Our technique for solving maximum weight stable set on a graph G given a tree decomposition of G is similar. The first step is to root the tree decomposition. A rooted tree decomposition of a graph G consists of a tree decomposition [T, W] of G and a rooted tree obtained by rooting T at some node r. For brevity's sake, we often use [(T, r), W] to denote this tree decomposition. It has the same width as [T, W]. Recall that Tt is the rooted subtree consisting of t and all its descendants. We use Gt to denote the subgraph of G induced by {v I v E WS for some s E Tt}. We use w(X) to denote >XEX w(x). The key to our dynamic programming algorithm is the following corollary of Fact 1.8.
Fact 2.3 For each node s in T, there are no edges between Gs - W3 and G - Gs. Furthermore, for any two children t and t' of s, there are no edges between Gt - Ws and Gt, - Ws. Now, in computing MWS for G given a rooted tree decomposition [(T, r), W]
of G, we actually compute for each node t of G and each stable set S in Wt, the value W (S, t) = max{w(S') I S' C V (Gt), S' stable, s = S' fl wt}. Then, the solution to this instance of maximum weight stable set is simply max{W(S,r) I S C W,., S stable}.
Brambles And Tree Decompositions
107
Now, if s is a leaf of T then for each stable set S in W8f W(S, s) = w(S). Fact 2.3 implies that if s has descendants t1, ... , t,, then for each stable set S in W8, we can choose a maximum weight stable set of Gs whose intersection with Ws is S by choosing a maximum weight stable set in each Gt, whose intersection with WS is S n Wt;. Thus, W(S, s) = w(S) + k
E max{W (Si, ti) - w(S n Si) I Si c Wt;, Si stable, S n Wt; = Si n W3}. i=1
If [T, W] has width w then for each node s there are at most 21+1 subsets S of Ws for which we may need to compute W(S, s). Furthermore using the equation above, we see that ifs has k descendants then we can compute W(S, S)
in O(k2w+1) time. We assume that [T, W] is a nice tree decomposition and hence has at most JV (G) nodes. It follows that our dynamic programming algorithm can be implemented in O(2w+ln) time which is linear for any fixed w.
We remark that at the cost of some extra bookkeeping but no increase in the computational complexity of the algorithm, we can easily find a maximum weight stable set instead of just the weight of such an object. Many other problems can be efficiently solved in graphs of bounded tree width by constructing a set of partial solutions corresponding to each node t of a bounded width tree decomposition. To do so, we need an efficient procedure for constructing the set of partial solutions corresponding to a node, given the sets corresponding to its descendants. Consider for example l Disjoint Rooted Paths for some fixed 1. Thus, we have a graph G, a rooted tree decomposition [(T, r), W] of G, and subsets X = {x1,. .. , xL}, Y = {yi,... , yL} of V (G).
Now, consider a set P = {P1,. .. , Pi} of paths in G where Pi links xi and yi. The restriction of P to G3 for some s in T is a set Q of paths of GS each element of which has its endpoints in (X n G8) U (Y n G3) U W. Thus Q has at most 1 + IW8 I < 1 + w + 1 elements. By a path scheme for s we mean a partition of (X n G8) U (Y n G3) U W8 into 1 + w + 1 or fewer ordered sets. We say a set {R1, . . . , R;} of disjoint paths in G8 is a realization of the path scheme 0 = 101,..., O; } if for each is Oi c_ Ri and the elements of Oi appear
along Ri in the given order. In this case we say that 0 is realizable. We note that the desired paths P 1 ,.. . , P` exist in G if and only if there is a realizable path scheme {Q,, ... , Qj} for r with j > l and {xi, yi} c Qi for each i between 1 and 1. To solve the given instance of l Disjoint Rooted Paths, we shall compute the realizable path schemes for each node s of T. Since IX U Y U W8 I < 21 + w + 1,
there are at most (21 + w + 1)21+w+1 path schemes for s. If s is a leaf we can determine which path schemes are realizable simply by considering the constant sized subgraph induced by W8. If s is not a leaf, then any realization of a path scheme for s corresponds to a set of edges within W8 and a set of realizations of path schemes for the descendants of s. This fact allows us to
108
B. A. Reed
compute all the realizable path schemes using dynamic programming in linear time if l and w are both fixed. In fact, with slightly more care we can solve l Disjoint Rooted Paths on graphs of bounded tree width in linear time even if 1 is part of the input.
Other well-known problems which can be solved in linear time given a bounded width tree decomposition of the input graph include: Clique, Hamilton Cycle, Chromatic Number, Domination Number, H-Minor Containment,
and Bandwidth. In fact it has been shown that any problem which can be formulated as a certain kind of logical formula [13, 6] can be solved in linear time on graphs of bounded tree width. All of the problems mentioned so far fall into this class. However, there are problems which can be solved in polynomial time on graphs of bounded tree width which do not fit this paradigm. We close this section with a discussion of two such problems (see [8, 7] for some more examples of problems which can be efficiently resolved on graphs of bounded tree width). Graph Isomorphism can also be solved in polynomial time on graphs of tree width at most w using dynamic programming, see [9]. However, the fastest algorithm known for this problem runs in 0(n'"+2) time. The algorithm considers two graphs Gl and G2 and a tree decomposition of Gl of width w.
For each node s of T, the algorithm computes which subgraphs of G2 are isomorphic to G. The fact that W3 contains at most w + 1 nodes allows us to consider at most 0(n'+') candidate subgraphs. We now show that the chromatic index of a graph G can be computed in linear time given a rooted tree decompsition [(T, r), W] of G. The argument we present is due to McDiarmid and Reed, it is almost ten years old although it appears for the first time here. Recall that the chromatic index of a graph G, denoted X'(G), is the minimum number of colours required to colour its edges so that any pair of edges which share an endpoint receive different colours. It is easy to see that the chromatic index of any graph G is at least its maximum degree which we denote A(G). Vizing [52] proved that in fact X'(G) < A(G) + 1. Thus to determine x'(G), we need only determine if the edges of G permit a A(G) colouring. A natural candidate for a partial solution corresponding to a node s of T, would be a A (= A(G)) colouring of those edges of G3 incident to W3 which extends to a A edge colouring of Gs. However since A may be n - 1, there can be exponentially many such colourings. Bodlaender [9] noted that it is sufficient to record the set of vertices of Ws incident to each colour in a A colouring of E(G,). He also noted that we do not need to record the names of the colours, it is sufficient to record for each subset of Ws, how many of the colour classes are incident precisely to this subset of W. He thereby restricted his attention to A2'+' = 0(n2'+') partial solutions at each node and derived a polynomial time algorithm. McDiarmid and Reed took a different approach. Reproducing work of Vizing, they showed that we can restrict our attention to a subgraph of G of
Brambles And Tree Decompositions
109
maximum degree less than 2w. In such a graph, there are at most 2w(w+1)2w 2w-edge colourings of the edges incident to any Ws and a linear-time algorithm follows immediately. Crucial to this approach is the following lemma due to Vizing.
Lemma 2.4 (The Adjacency Lemma) Let uv be an edge of a graph F of maximum degree at most k such that d(u) + I{w I w E N(v), d(w) = k}I < k. Then X'(F) < k if and only if X'(F - uv) < k. (This lemma is vacuously true unless k = 0(F) in which case it is equivalent to: either O(F-uv) = A(F) -1 in which case X'(F) = A(F), or X'(F) = X'(F - uv).)
By a standard reduction of F we mean a graph obtained by repeatedly removing edges satisfying the conditions of the adjacency lemma for k = A(F).
It is not difficult to see that there is in fact a unique standard reduction of any graph F, and that this graph either has the same maximum degree as F or has no edges. Now, the colouring number of F, denoted 6*(F), is the maximum over all the subgraphs of F of the minimum degree. Fact 1.10 implies that d*(F) is at most the tree width of F. Applying the adjacency lemma, we obtain:
Lemma 2.5 If A(F) > 2b*(F) then the standard reduction of F is empty and hence X'(F) = O(F). Proof Consider the standard reduction H of F. If H has no edges then X'(F) = 0(F) and we are done. So, to prove the lemma, we need only show that if H has edges then O(F) < 2J* (F). To this end, assume that H contains edges and hence O(H) = O(F). Consider the set D of vertices of degree
at least O(F)/2 in H. Let v be any vertex of D. If v has a neighbour in V(H) - D then by the adjacency lemma and our choice of H, v must have 0(F)/2 neighbours of degree O(F) in H. Thus, every such v has at least z/2 neighbours in D. So the vertices of degree at least O(F)/2 in H span a subgraph of minimum degree at least O(F)/2. The result follows. McDiarmid and Reed noted that it is easy to compute the standard reduction G' of our graph G of tree width at most w using dynamic programming in linear time. If G' has no edges then the chromatic index of G is A, and our algorithm has no work to do. Otherwise, by Lemma 2.5 and the fact that S* (G) is at most w, we obtain that A is at most 2w, and as remarked above dynamic programming can be used to compute X'(G) in linear time. To actually find an optimal colouring, we need to repeatedly rip out matchings which decrease the chromatic index. Such a matching can be found in linear time using dynamic programming. This yields a quadratic algorithm for edge colouring graphs of bounded tree width. Taking more care, we can obtain an algorithm which runs in 0(n log n) time. Can you find a linear time algorithm for this problem or a more efficient algorithm for Isomorphism Testing on graphs of bounded tree width?
B. A. Reed
110
Finding tree decompositions of bounded width
2.2
In the last section, we showed how various optimization problems could be solved efficiently in a graph, given a tree decomposition of the graph with bounded width. In this section, we describe, for each fixed k, an 0(n2) algorithm which given a graph G, either determines that G has tree width greater than k, or finds a tree decomposition of G of width at most 4k + 1. To describe the algorithm we need to introduce the notion of separators. For our purposes, a separator for a graph G is a subset X of V (G) such that no component of G - X contains more than 31 V (G) - X I vertices. An S-separator for some S C V(G) is a set X of vertices of G such that no component of G-X contains more than 3 1S - X vertices of S. The following fact links the notions of separators and tree decompositions:
Fact 2.6 If G has tree width at most k then for all S C V (G), G has an S-separator of order at most k + 1.
Proof Obviously every set S of at most k + 1 vertices is itself an S-separator with at most k + 1 vertices. So we consider only subsets of the vertices with order at least k + 2. To begin, we define for each set X in V (G), the set ADJ(X) which is {v I v ¢ X and v is adjacent to some vertex in X}. Now, for any set S of at least k + 2 vertices of G, we define a set ,QS of connected subgraphs of G as follows. A connected subgraph B of G is in Ns if and only if both I ADJ(B) I < k + 1 and IB n SI > 3IS - ADJ(B)I. We will show that either G has an S-separator of order at most k + 1 or /3s is a bramble of order at least k + 2 and hence the tree width of G is at least k + 1. We show first that Ns is a bramble. To this end consider elements B1 and B2
of I3S. If B1 and B2 do not touch then B1 C_ G - ADJ(B1) - B2 - ADJ(B2)
and B2 C_ G - ADJ(B2) - B1 - ADJ(B1). Thus, IB1 n SI + IB2 n SI
31S-ADJ(Bi)l and IB2nSI >
31S -
ADJ(B2)I. So IB1 n SI + IB2 n SI > IS - ADJ(B1) - ADJ(B2)1. This contradiction shows that B1 and B2 touch and hence Ps is a bramble. If H is a hitting set for 13s of order at most k + 1 then no component of G - H contains more than IS - HI vertices of S, for such a component would 3 be in ,Qs, a contradiction. Thus, H is an S-separator. Otherwise, ,QS has order at least k + 2. This completes the proof of Fact 2.6.
The algorithm we describe actually provides an algorithmic proof of:
Fact 2.7 If G has an S-separator of order at most k + 1 for all S C V (G) then G has tree width at most 4k + 1. Combining these two facts, we see that we can determine the tree width of a graph to within essentially a factor of four simply by finding a smallest
S-separator for each S C_ V(G). Thus the notion of tree width is indeed intimately linked to that of separators.
Brambles And Tree Decompositions
111
Remark Planar separators have received much attention in the literature. A fundamental result, due to Lipton and Tarjan [26] is that every planar graph has a separator containing at most 2 2 IV (G) I vertices. Alon, Seymour, and Thomas [2] improved the bound to 2 2 IV (G) vertices by studying an appropriate tangle. They also bounded the size of a minimum separator in a graph with no Ki minor for a fixed 1, see [1].
Remark We note that if instead of the value 3 IS - X I in the definition of separator we used 21SI then a set S would have no separator of order less than k if and only if it were k-linked. The fraction 3 is used for laminarity reasons, just as we often use tangles instead of brambles. We could prove similar results using the fraction 2. This relationship between separators, a well studied notion [12], and both linkedness and bramble number, deserves further study.
Furthermore, we recall that, as we saw in the proof of Fact 2.6, a set S without a separator of order k defines a bramble of order k. So, Fact 2.7 also implies that the tree width of G exceeds its bramble number by at most a factor of four. We sharpen this result in the next section. Our algorithm for finding tree decompositions recursively uses separators to decompose the graph into subgraphs and then pastes together tree decompositions of these subgraphs. In order to do the pasting we need the following, which the reader may find more intelligible if he is looking at Figure 9. Fact 2.8 Let X1 and X2 be two sets of vertices in a graph G with X1 c X2.
Let C1, ..., C1 be the components of G - X1. For i between 1 and 1, let Y = (X2 fl Ci) U X1, let [Ti, W'] be a tree decomposition of Ci U X1. Suppose that for each T' there is a node ti of Ti with Y C Wt,. Let T be a tree obtained from U1_1 Ti by adding a node t adjacent to each of tl, ..., ti. For s E Ti, let
WS = W. Let Wt = X2, and let W = (W3 I s E T). Then IT, W] is a tree decomposition of G.
Proof Each edge xy of G appears in some X1 U Ci so {x, y} C WS = W3 for
some sinTi. For each v in C1 - X2, {s I v is in W3} = {s I v is in W} is a subtree of T because it is a subtree of Ti. For each v in Ci fl (X2 - X1), we have Is I v is in WS} = Is I v is in Ws} + t is a subtree of T because ti is in the tree Is I v E Ws} and t is adjacent to ti. For each v in X1i Is I v E Ws} is the subtree of T consisting of t and for each i the subtree Is I v is in W, 'J of Ti (which contains ti).
Remark We actually need only construct [T', W'] for i with Ci ¢ X2 as for any Ci C X2, Ci U X1 is contained in Wt.
Without further ado, we present the algorithm.
B. A. Reed
112
G
Wt; 2Xlu(X2nCi) [T, W]
Figure 9:
Brambles And Tree Decompositions
113
Algorithm: k-Tree Finder
Input: A graph G = (V, E) and a subset W of V with I W Output: Either,
3k + 1,
(i) a tree decomposition [T, W] of G of width at most 4k + 1 such that W C Wt for some t E T, or
(ii) a subset S of G such that G has no S-separator of order k + 1 or less.
Description of Algorithm If G has 4k + 2 or fewer vertices then return a tree decomposition for G which uses a one node tree. Otherwise, arbitrarily add vertices to W until IW I = 3k + 1. Next, attempt to find a W-separator X of order at most k+1 (using a procedure described below). If no such separator exists, return with output (ii) W. Otherwise, let U1, ..., U1 be the components
of G - X which are not contained in W. Let Gi = X U Ui and let Wi = (Ui n W) U X. Note that for distinct i and j, Gi n G; C X. Since X is a W-separator, IW I < 3k + 1, and JX I < k + 1, it follows that for each i, I Wi I < 3k + 1. Thus, we can apply k-Tree Finder to (Gi, Wi) for each i. If it turns out that for some i there is an Si such that Gi has no Si-separator of order at most k + 1 then clearly G also has no Si-separator of order at most k + 1 so we return (ii) Si and stop. Otherwise, we find for each i, a tree decomposition [Ti, Wi] of Gi (we let Wt be the element of Wi corresponding tot E Ti) and a distinguished node ti of Ti such that Wi C Wt.. In this case, we obtain a tree decomposition [T, W] of G by setting:
(a) V (T) = UV(T) U {t}, (b) E(T) = U E(Ti) ti=1
U
(tJ{})
and
(iOlfi-ti}
(c) W = U Wi U (X U W) (that is, Wt = X U W and W9 = WS for i=1
sETi).
It follows from Fact 2.8 that [T, W] is indeed a tree decomposition. It is easy to verify that it has width at most 4k + 1 so we return (i) ([T, W], t). Now, the key step in this recursive procedure is finding a W-separator of order at most k + 1. We then separate our problem into l smaller practically disjoint subproblems. As we shall see, this implies that we consider at most 0(n) subproblems. In fact, we shall prove:
114
B. A. Reed
Claim 2.9 Whilst applying k-Tree Finder to a graph G, we apply the algorithm to at most max(1, 21V(G) I - 6k - 3) subproblems (including the original one). Hence, if we return a tree decomposition then the associated tree has at most max(1, 21V(G)I - 6k - 3) nodes. We now consider the time necessary, during a particular application of the algorithm, to split the problem into subproblems and then combine the solutions to these subproblems. We shall show that we can do this in O(IE(G)I) time. Clearly, it then follows from Claim 2.9 that the algorithm runs in O(IV(G)IIE(G)I) time. Note that the difficult part of splitting the problem into subproblems is determining if G has a small W-separator and finding one if it does. If this can be done in 0(IE(G)I) time then so can the whole splitting process. Furthermore, building IT, W] from {[T1, W1], ... , [T1, W`]} takes O(ITI) time which by Claim 2.9 is 0(n). So, the fact that the algorithm runs in 0(IE(G)IIV(G)I) time follows from:
Claim 2.10 Given a set S of at most 3k + 1 vertices in a graph G, we can determine if G has an S-separator with k + 1 or fewer vertices in 0(k * 33k+1 * IE(G)I) time.
To reduce the time complexity to 0(n2) we add a preprocessing step which counts the number of edges of G. If G has more than kn edges then by (1.10) its tree width is at least k + 1. We could just return with this fact and stop. If we want to return a set S with no (k + 1)-separator then we can run k-Tree Finder on a subgraph G' of G consisting of some kn + 1 of its edges. To complete our analysis of the algorithm it remains only to prove the two claims.
Proof of Claim 2.9 We prove the claim by induction on IV(G) I. If IV(G) I < 4k + 2 then we make no recursive calls to the algorithm and the bound trivially holds. So, we assume that IV (G) I is at least 4k + 3 and the bound holds for all graphs with fewer vertices than G. Now, if the algorithm fails to find a Wseparator then again we make no recursive calls and the bound trivially holds. Otherwise, we find a W-separator X with k+1 vertices and create proper subgraphs G1i ..., G1 of G for some 1 > 1, and apply k-Tree Finder to each of these subgraphs in turn. Clearly, if ni is the number of subproblems considered when we apply k-Tree Finder to (Gi, Wi) then we consider 1 + E1=1 ni subproblems throughout our application of k-Tree Finder to (G, W). By the induction hypothesis, ni < max(1, 2I V (Gi) I - 6k - 3). If l = 1 then since X is a W-separator and I W I = 3k + 1, we have that I V (Gi) I < IV(G) 1. The claim follows immedi-
ately by induction. So, we can assume that l > 2. Note that for any distinct i and j, V(Gi)-Wi and V(G3)-Wj are disjoint non-empty subsets of V(G)-W. Furthermore, ni < max(1, 2IV(Gi)I - 6k - 3) < 2IV(Gi) - Wil - 1. It follows that 1 + Ei=1 ni < 1 + 21V(G) - WI - 1 < 2IV (G) I - 6k - 3, as required.
Brambles And Tree Decompositions
115
Proof of Claim 2.10 Consider a set S of at most 3k+1 vertices of a graph G. First, we note that G has an S-separator with k + 1 or fewer vertices if and only if there is a partition of V (G) into A, B, and X such that the vertices of each component of G - X are contained either in A or in B, IX I < k + 1, I A n S I< IS - X I, and IB n SI < s IS - X I. (Obviously, if such a partition exists then3X is an S-separator. Conversely, given an S-separator X we can order the components of G - X as U1, U2, ..., U1 so that IUD n SI > IUz+1 n SI
and then obtain our partition by setting A = Ui_1 U, where j is the minimal integer for which I S n U; I > IS - X1.) So, we will attempt to find such a partition rather than looking for san S-separator directly. U2+1
We do this by determining separately for each of the 33k+1 choices of three
disjoint sets SA, SB, Sc whose union is S whether or not there is such a partition of G with SA = A n S, SB = B n S, and Sc = X n S. If there is no choice of SA, SB, and Sc for which the desired corresponding partition of G exists then obviously G has no S-separator of order at most k + 1. If for some choice of SA, SB, and SC there is a corresponding partition of G then we will find an S-separator X with X n S = Sc when considering this partition. So, we turn our attention to a particular choice of SA, SB, and Sc. We first ensure that I SA I C 2I SB I, I SB I < 2I SA I and ISCI < k+1. We then check in 0(kIE(G)I) time, using standard alternating path techniques, whether there are k + 2 - ISCI internally vertex disjoint paths from SA to SB in G - Sc. If there is no such set of paths then we find a set X of k + 1 vertices with
X n S = Sc such that there are no SA to SB paths in G - X. Letting A be the union of those components of G - X which contain an element of SA and
letting B = V (G) - A - X yields a partition of G which shows that X is an S-separator. Conversely, if there is a partition of G corresponding to this choice of SA, SB and SC then there cannot be k + 2 - I SCI internally vertex disjoint SA to SB paths in G - SC so we will find some S-separator.
This completes our proof that k-Tree Finder runs in 0(n2) time. Now, Robertson and Seymour originally introduced k-Tree Finder in [36]. The implementation given in that paper has running time bounded by a polynomial whose exponent is a fast growing function of k. In [42] they presented the current version. Reed [32] by introducing further technical complications, speeded
up the algorithm. His version runs in 0(nlog(n)) time. We remark that although these algorithms give bounded width tree decompositions of graphs of tree width at most k, they do not determine the tree width of the graph exactly.
In [36], Robertson and Seymour pointed out that since the class of graphs with tree width at most k is minor-closed (this is left as an exercise for the reader), they could test if a graph has tree width exactly k by testing for Hminor containment for each graph H in the obstruction set for this class. They noted that given the tree decomposition of width at most 4k + 1 returned by k-Tree Finder they could solve each such problem in linear time. They proved,
B. A. Reed
116
by applying a special case of (1.14) that they had already resolved, that this obstruction set was finite. It follows that they can determine, for a fixed k, if a graph has tree width at most k by first applying k-Tree Finder and then applying a linear-time algorithm. Note since the obstruction set for graphs of tree width at most k is not known, they proved only the existence of an algorithm. They gave no way to construct one. In 1987, Arnborg, Corneil, and Proskurowski [4] independently constructed an O(n'+2) algorithm for determining, for fixed k, if G has tree width at most k and constructing a width k decomposition if one exists. They also showed that determining if G has tree width k is NP-complete if k is part of the input. In 1993, Arnborg et al. [5] gave a linear time algorithm to determine if the tree width of a graph is exactly k, for k fixed. However, this algorithm required much more than linear space (reading in unwritten memory is permitted) and if we actually wanted to find the tree decomposition O(IV(G) 12) time would be required. Bodlaender and Kloks [11] developed a straightforward method for testing the tree width of a graph given
a bounded width tree decomposition of it. This method runs in linear time, and actually constructs an optimal tree decomposition. Bodlaender [10], by combining this algorithm with some novel techniques developed a linear time algorithm to determine, for fixed k, if G has tree width k and to construct a tree decomposition of width at most k if one exists. 2.3
A duality theorem
Recall that a bramble is a set of connected subgraphs each two of which touch, that is intersect or contain distinct endpoints of some edge. A set H of vertices in G is a hitting set for a set of subgraphs of G, if each of the subgraphs intersects H. The order of a bramble B is the minimum of the orders of the hitting sets for B. The bramble number of G, denoted BN(G), is the maximum of the orders of its brambles. In this section, we prove: Theorem 2.11 ([48]) The tree width of G is exactly one less than its bramble number.
Now, as we saw in the introductory section on tree decompositions (see Fact 1.4), for any tree decomposition [T, W] of G and any bramble B in G, there is a t E T such that Wt is a hitting set for B. Thus, TW(G) > BN(G) -1; consider a maximum order bramble and a minimum width tree decomposition. It remains only to prove the converse. We actually prove a stronger statement which characterizes when a bramble can be extended to a bramble with order at least a given 0. To wit,
Lemma 2.12 For any bramble B in G, either (i) there is a bramble B' of order 9 with B C 13', or
Brambles And Tree Decompositions
117
(ii) there is a tree decomposition [T, W] for G such that if t is a node of T with IWtI > B then t is a leaf and Wt is not a hitting set for B.
We remark that this lemma implies that TW(G) < BN(G) - 1; consider a bramble with no elements. We also observe that (1.4) implies that at most one of the two possibilities discussed in Lemma 2.12 can occur.
Proof of Lemma 2.12 Consider a graph G. We shall show that one of (i) or (ii) holds for each bramble B in G by assuming the contrary and deriving a contradiction. So, we choose a bramble B which satisfies neither (i) nor (ii) and which has fewer hitting sets with at most 0 - 1 vertices than any other such counterexample to the lemma. If there are no such small hitting sets for B then it is a bramble of order 0 and thus (i) holds for B, a contradiction. So, we let H be a minimal order hitting set for B and note that I H I < 8 - 1. The idea of the proof is quite simple. Note first that H V (G) as otherwise there is a tree decomposition of G satisfying (ii) which uses a one node tree. So, we let C1, ..., Ct be the components of G - H. For each Ci, we shall find a tree decomposition [Ti, Wi] of H U Ci satisfying:
(a) if t is a node of Ti with W. > 0 then t is a leaf and
(*)
Wt is not a hitting set for B, and (b) there is a leaf ti of Ti with Wt: = H. We then form a tree decomposition [T, W] of G by taking a copy of each of these tree decompositions, adding a node t adjacent to {t1, ... , tt}, and setting Wt = H, as discussed in Fact 2.8. Now, each [Ti, Wi] satisfies (*)(a) so [T, W] shows that condition (ii) in the lemma holds for B, a contradiction. Thus to prove Lemma 2.12 and the theorem we need only show that for each component Ci of H, there is a tree decomposition [Ti, Wi] of H U Ci satisfying (*). In doing so, we consider two possibilities. The first is that Ci fails to touch some element B of B. In this case, let Ti be a tree with one
edge st, let Ws = Ci U H - B, let Wt = H and let Wi = {W37 Wt}. Since Ci does not touch B, [Ti, Wi] is a tree decomposition of H U Ci. Clearly, it satisfies (*). The second possibility is that Ci touches every element of B and thus B+Ci
is a bramble. In this case we remark that, since H is a hitting set for B but not B+Ci, by our choice of B, one of (i) or (ii) holds for B+Ci. If (i) holds for B + Ci then it also holds for B, a contradiction. Thus, (ii) holds for B + Ci, so we consider a tree decomposition [T', W'] of G such that for every node t of T' with I Wt l > 8, t is a leaf and Wt is not a hitting set for 8 + Ci. Now, we can
assume that for some leaf t in T', Wt fails to intersect Ci but is a hitting set for B, as otherwise [T', W'] would show that B satisfies (ii), a contradiction. We will show that we can transform [T', W'] into a tree decomposition of HUCi satisfying (*). To do so, we need the following lemmas which we prove in a moment.
B. A. Reed
118
Lemma 2.13 Let L be a bramble of order k in a graph F. Let H1 and H2 be hitting sets for L. Then there are k vertex disjoint paths between H1 and H2 in F.
Lemma 2.14 Let IT, W'] be a tree decomposition of a graph G. Let
X be a cutset of G and let C be a component of G - X. Suppose there is some r in T such that W,: is contained in G - C and there are IX I vertex disjoint paths of G between X and Wr. Then there is a tree decomposition IT, W] of X + C such that: (i) for each s in T, I Ws I < 1W' 1, and
(ii) for each leafs of T except r, W3 C Ws.
Now, B clearly has order JHI and both H and Wt are hitting sets for B. Thus, Lemma 2.13 implies that there are 1171 vertex disjoint paths between H and Wt. So, we can apply Lemma 2.14 to the tree decomposition [T', W']
with X = H, C = Ci, and r = t. This yields the desired tree decomposition IT, W] of H U Ci satisfying (*) (to see this note that for any leaf s of T' with IW:l > 0, either Ws is not a hitting set for B, or Ws does not intersect Ci in which case W9 is contained in H and hence has at most 0 - 1 elements). Thus, to complete the proof of Lemma 2.12 and our theorem we need only prove Lemmas 2.13 and 2.14.
Proof of Lemma 2.13: Let F, G, k, H1, and H2 be as in the statement of the lemma. Note that every element of L intersects both H1 and H2 and therefore contains a path from H1 to H2. Thus, if X is a set of vertices of G such that there is no path between H1 - X and H2 - X in G - X then X is a hitting set for L. So, any such X contains at least k vertices and the lemma follows by Menger's theorem.
Proof of Lemma 2.14 Let G, IT, W'], X, C, t, and W,: be as in the statement of the lemma. Let k = IXJ and let P1, ..., Pk be k vertex disjoint paths from X to W. Let xi be the element of X contained in Pi.
Note that since X is a cutset, C is a component of G - X, and none of W,: is in C we know that no Pi intersects C. Now, we know that Spi = U{S I v E V(Pi)} is a subtree of T containing Sri and r. Thus there is a path Qi of Sp, with one endpoint r and the other in S. Since Qi and Sri are intersecting subtrees of T, Sri U Qi is a subtree of T. Now, we obtain our tree decomposition IT, W] of X + C by setting, for each s in T, Ws = (WS fl (X + C)) U {xi I s is on Qi}. To see that we do indeed obtain a tree decomposition note first that for each vertex v of C, W']) and so is a subtree of T. Furthermore, for Is I v is in W3} = v in X we have that Is I v is in WS } = S ([T, W']) U Qi and hence is a subtree of T. Finally, for any edge e = xy of X + C, there is some s such that Ws contains {x, y} and therefore so does W.
Brambles And Tree Decompositions
119
We note that for each leaf s of T apart from r if Qi contains s then s is the endpoint of Qi in Sam; ([T, W']) and hence Ws contains xi. Thus,
W3 c W. Furthermore, if s is a node of T then
WS-X cw nCcww-Pl-P2-..._pi and
I {xi I xi E Ws n X} I
_ I{i I Pi intersects Ws} I Ify yis on some Pi,yEWW}1.
Thus IW91 < 1 Ws as claimed. This shows that [T, W] is a tree decomposition with the properties we desire. That completes the proof of Lemma 2.12.
3
Untangling tangles
We have two objectives in this section. One is to gain a deeper understanding of tangles, in particular which of their properties make them better behaved than other brambles. First however we will prove (1.13) which we restate below for the reader's convenience.
The Canonical Tree Decomposition Theorem Given a graph G, we can construct a tree decomposition T which has the following properties.
(1) For each maximal tangle T of G, there is exactly one node t(T) of T whose vertices form a hitting set for T.
(2) If Ti and 'Ti are indistinguishable maximal tangles (and hence have the same order) then t(Ti) = t(T ).
(3) If Ti and 'Ti are distinguishable maximal tangles then t(T1) t(T2) and there is an arc st on the unique t(Ti) to t(T) path of T such that -`
V (X,) n V (Xt) is a (Ti, T )-distinguisher of minimum order.
(4) For every node t of the tree decomposition, there is a maximal tangle T
such that t(T)= t. Proof Our first step in constructing this tree decomposition is to define an equivalence relation - on the set of maximal tangles of G such that Ti ' ' 'Ti precisely if Ti and T are indistinguishable (and hence by the definition of maximal, ord(T1) = ord(T)). To simplify matters, we want to choose a representative tangle from each equivalence class of c. To this end, note that for any tangle T, of(-r) is a tangle of the same order as T which is indistinguishable from T; in fact, of (T) = {Y I Y is a component of G - X for some X with fewer than ord(T) vertices and Y contains an element of T}. So, we define a
B. A. Reed
120
tangle T to be canonical if each element Y of T is a component of G - X for Of(T) some X with JXI < ord(T). Obviously, if T is canonical then = T. Furthermore, if Ti ^ T then Of (-r,) = Of (-r,). Thus, we see that for any tangle T, #f(T) is the unique canonical tangle in the equivalence class of - containing T. We let TG be the set of canonical maximal tangles. By the above remarks, TG contains exactly one representative from each equivalence class of -. We are going to choose for each pair of tangles Ti and Tz in TG a minimal (T1, T )-distinguisher and then use these distinguishers to construct the desired tree decomposition. To do so, we recall that V (G) = {v1,. .. , v,,} and E(G) _ are indexed sets. lei, . . . , If T1 and Tz are two tangles of G and the ordered pair (A, B) is a separation of G whose order is less than min(ord(T1), ord(T)) and one of the two graphs A - X or B - X completely contains an element of T1 whilst the other completely contains an element of T then we say that (A, B) is a (T, Tz)separation. Note that if (A, B) is a (T1, T2)-separation then V (A) n V(B) is a (Ti, T )-distinguisher. Conversely, it is easy to see that if X is a (T1, T)distinguisher then there is at least one (Ti, T2)-separation (A, B) with X = V (A) n V(B), we simply need to assign each component of G - X to one of A or B so that the two components which completely contain an element of one of the two tangles are assigned to different sides of the separation. The canonical (T1, Tz)-separation for two distinguishable Ti and T2 is defined to be the unique (T1, T2)-separation (A, B) which:
(i) has minimum order,
(ii) subject to (i), lexicographically minimizes V(A) fl V(B) (where S is lexicographically less than R if the lowest indexed element of S U R which is not in S fl R is in S), (iii) subject to (ii), lexicographically minimizes V(A),
(iv) subject to (iii), maximizes (E(A)I (this is equivalent to requiring that all of the edges of G with both endpoints in V(A) fl V(B) are in A.)
We let SG be the set of precisely those separations which are canonical (Ti, T2)-separations for some pair (Ti, T) of tangles in TG. We shall show that SG is a set of laminar separations. As noted in the introduction, this implies that there is a unique tree decompostion [Ts0, Xs0] of G such that each separation S in SG is made by precisely one arc of TSG and for each arc a of TsG one of the separations made by a is in SG. To complete the construction we have been describing, we simply need to show that the tree decomposition [TSG, XSG] satisfies properties (1)-(4). Forthwith the details.
Lemma 3.1 For every graph G, the set SG is laminar.
Brambles And Tree Decompositions
121
C
AnC-B1T-D''
AV nBV
nCv-D`'
BnC-A"'-Dv
r A
CvnDvnAV - B''
AnD-BV-Cv
AV nBV
nCvnDv AvnBv
nD"-Cv
CvnDvnW'-A''
B
BnD - AV - C''
D Figure 10: Here F" denotes V(F)
Proof Let (A, B) be a (Ti,Tj)-separation of minimum order and let (C, D) be a (Tk, T1)-separation of minimum order (see Figure 10). Choose
the labels so that (where (i) the letters in the next equation and in the right hand sides of the following three equations correspond to vertex sets and (ii) V(F n H) = V(F) n V(H)): 1(C n D) n (A - B) I
h. Hence, X31 < h and ,33 = fT(X3) exists. Now, note that X3 also separates Z from T. For, we know that either 01 n 02n 03 is non-empty or there is an edge e with an endpoint in each of 01, 02, and 03. However, in the latter case, by the definition of X3, the endpoint of e in 03 must be in ,Ql n/2 n,Q3. So, 01n,32 na3 is non-empty and, again by the definition of X3, we obtain that /33 C 01 n,32. So, indeed /33 does not intersect Z. But now, X3 contradicts either the fact that Z is T-linked or our choice of X1 and X2. This contradiction proves the existence of the desired X*. Now,
assume that X* is not T-linked. In this case, there is some T in T and a set Y of at most h - 1 vertices separating X* from T. Note that T is disjoint from X* and thus must be contained in fy(X *). So there is no path in G - X from Z to T and hence Y also separates Z from T. But this contradicts the fact that Z is T-linked. So, we see that X* is T-linked. It follows that there are h vertex disjoint paths from Z to X* as both these sets are T-linked. We note that just as with the analogue of (3.1), the closest cutter lemma fails to hold for the brambles in a cycle. To see this consider a cycle of length four with vertices {v1, v2, v3, v4}, appearing in that cyclic order around the cycle. Now, ,6 = {{v1}, {v2}, {v3, v4}} is a bramble in this graph. Furthermore, X = {v3, v4} is a 3-linked set. However, both {v4, v2} and {v1, v3} are cutters for X. This shows that the closest cutter lemma does not hold for this bramble. We now prove two lemmas about T-linked sets. The first requires us to define a new connectivity invariant tied to the bramble number. Definition We say a set X of vertices is well-linked if for every pair A and B of subsets of X with JAI = IBI there are CAI vertex disjoint paths between A and B. We define the well-linkedness of G, denoted WL(G), to be the size of the largest well-linked set in G.
Remark Note that we do not insist that A and B are disjoint thus if we required the paths to have no internal vertices in X we would have an equivalent definition.
Lemma 3.4 BN(G) < WL(G) < 4 BN(G).
Brambles And Tree Decompositions
127
Proof To prove the first inequality, we show that any minimum hitting set H for a bramble a is a well-linked set. Otherwise, by Menger's Theorem, there are two equal sized subsets A and B of H and a set Y of fewer than JAI ver-
tices separating A and B. By symmetry, we can assume that A n fp(Y) is empty. Now, we know that Y U (H n fp(Y)) is a hitting set for /3 contained in (H - A) U Y. However, this contradicts the minimality of H.
To prove the second inequality, we show that any well-linked set X is (kx)-linked where kx = FIX /41. Since every k-linked set defines a bramble of order k, the result follows. So, assume for a contradiction that there is a set W of fewer than kX vertices such that no component of G - W contains more
than half the vertices of X. This implies that we can partition G - W into two sets C and D each of which contains more than a quarter of the vertices of X. (To see this, let U be the component of G - W which has maximum intersection with X. Set C = U and recursively add more components to C until C contains more than a quarter of the vertices of X.) Now, choose a subset A of kx vertices of X n C and a subset B of D n X of the same size. The sets A, B, and W contradict our assumption that X is well-linked. Lemma 3.5 Let ,Q be a bramble. Every 0-linked set is well-linked.
Proof We mimic the first half of the proof of the last lemma to show that if X is not well-linked and has cardinality less than ord(,3) then there is a subset A of X and a set Y of smaller cardinality than A such that A is disjoint from f# (Y). Then YU (X n f# (Y)) separates X from the hitting set YU f# (Y). Thus X is not /3-linked.
Lemma 3.6 If T is a tangle then every T-linked set extends to a T-linked set of order ord(T) - 1.
Proof Let X be a T-linked set of order at most ord(T) - 2. Let X* be the closest cutter for X. Let y be any element of fT(X*). We claim that X +y is a T-linked set. Otherwise, there would be a set Z of fewer than IX + yJ vertices
separating X + y from T. Since X is T-linked, we know IZI = IXI. Since X* is the closest cutter for X, we know fy(X*) C fr(Z). Thus y E fr(Z) contradicting the fact that Z separates X + y from T. Finally, to close this section, we present two more definitions.
Definition Let Ti be a tangle in a graph G, and let T be a tangle in a subgraph of G. Obviously T is also a tangle in G. We say T is conformal with Ti if the order of Ti is at least that of T, and they are indistinguishable in G.
Definition Let T be a tangle in a graph G. Let Z be a set of k < ord(T) vertices of G. Define T/Z to be IT E T 1 T n Z = 0}. Then T/Z is a tangle of order at least ord(T) - k in G - Z. Furthermore, it is conformal with T.
B. A. Reed
128
4
Excluding walls
In this section, we give a brief sketch of the proof of Theorem 1.2, which states that if a graph G has no wall of height h then it has no bramble of order greater than 2534h5. The proof of this theorem is long and fairly technical; we delay discussing its details until Section 7. We will however discuss it briefly in order to present a strengthening which we will need in the next two sections.
To begin, we recall that if a graph has a bramble of order 3k then it has a tangle of order k. The main idea of the proof is to start with a large T-linked set for some huge order tangle T and then grow a set P of paths from this set "towards" the tangle. We use a subset of these paths to form the rows of our wall. We shall show that we can also find a set C of disjoint paths between the appropriate elements of P which we use to form the columns. In order to construct the paths of C we will need to repeatedly use a lemma which shows that given two vertices s and t in a T-linked set X we can find a path P between s and t in G - (X - s - t) and a tangle T' in G - P whose order is not much smaller than that of T and such that X is a T'-linked set. The precise lemma we need is (see also Figure 12):
H
G' Figure 12:
Lemma 4.1 ([46]) Let h and B be integers with 0 > h > 2. Let T be a tangle in a graph G of order at least 24(h + B) + 7, and let Z be a T-linked set of h vertices of G. Let s and t be two vertices of Z. Then, there are subgraphs H and G' of G such that setting Z' = V(H) n V (G) we have:
( i ) IZ'J = h and there are h vertex-disjoint paths P1, ... , P1 in H from Z
to Z' and an s to t path Q in H - Z' such that for each i, Q n Pi is a (possibly empty) path, and
Brambles And Tree Decompositions
129
(ii) there is a tangle T' in G' of order 0 and conformal with T such that Z' is a T'-linked set. Repeated applications of this lemma and some heavy slogging allow us to
build up a column C1 from various paths like Q. By iterating the column building procedure, we can knit together a wall W of height h, adding the columns one by one. The reader may find that Figure 13 aids his intuition. Note that the set of rows is obtained by concatenating a family of sets of subpaths which are obtained via Lemma 4.1. We remark that both the set of the initial endpoints and the set of final endpoints of each set in this family is a T-linked set. Now, for each i > 1, we let Hi be the connected subgraph consisting of Ci and for each row R, the portion of R3 strictly between Rjf1Ci_1
and Rj fl Ci. Our remark implies that each Hi contains a T-linked set Zi containing h vertices one on each row. This fact can be used to obtain the strengthening of Theorem 1.2 stated as Theorem 4.2 below.
zo
Z1
Z2
Z3
Figure 13:
Theorem 1.2 states that given a large tangle T in a graph G, we can use T to find a high wall W. We actually show that we can find such a wall which is not separated from T by any small order cutset. To be more precise, we need a definition. We say a wall W of height h is attached to a tangle T if T has order at least h and for every set X of fewer than h vertices fy(X) is the unique component of G - X containing a row (and hence a column) of W. We can extend Theorem 1.2 to prove:
Theorem 4.2 Let h be an integer. Let T be a tangle of order at least in a graph G. Then there is a wall W of height h attached to T.
2534h5-1
Proof Idea Construct the wall W as discussed above. Let X be any set of at most h vertices of G. Clearly, X fails to intersect at least one of the Hi defined above, and hence there is some i such that Zi is in a component U of G - X. Now, as Zi is a T-linked set, U must be fy(X). X also fails to intersect some row Rj. Thus, since R; intersects Zi, Rj is contained in fy(X). So, fy(X) is indeed the component of G - X completely containing a row of W.
B. A. Reed
130
5
Graph minors revisited
We remind the reader that two of the most important results concerning graph minors proved by Robertson and Seymour were:
(1.14) Wagner's Conjecture In any infinite sequence G1, G2, ... of graphs, there exist i # j such that Gi is a minor of Gi.
(1.16) There is a polynomial-time algorithm to resolve k Disjoint Rooted Paths for any fixed k (Robertson and Seymour's algorithm runs in O(I V (G)13) time, Reed improved this to O(I V (G)12).)
We now investigate the role that tree decompositions played in the proof of these results. We consider (1.16) first and we begin with a definition.
Definition Let (G, X, Y) be an instance of k Disjoint Rooted Paths. A vertex v is irrelevant (with respect to (G, X, Y)) if the desired paths exist in G if and only if they exist in G - v. Now, as mentioned in Section 2, Disjoint Rooted Paths for fixed k is easy to solve in linear time on graphs of bounded tree width. Robertson and Seymour [42] proved:
Theorem 5.1 For every k there is an hk such that if (G, X, Y) is an instance
of k Disjoint Rooted Paths and W is a wall of height hk in G then there is an irrelevant vertex v in W. Furthermore, such a wall and corresponding irrelevant vertex can be found in polynomial time. (Robertson and Seymour's algorithm runs in O(IV(G)12) time, Reed improved this to O(IE(G)I)). Now, obviously having found an irrelevant vertex v for (G, X, Y) we can
restrict our attention to (G - v, X, Y). Robertson and Seymour repeatedly apply Theorem 5.1 and delete the irrelevant vertex it returns until the graph they are considering contains no high wall. Theorems 1.1, 1.2, and 1.3 imply that such a graph has tree width at most 25sank and hence we can solve the k Disjoint Rooted Paths problem using dynamic programming. We now briefly sketch the methods Robertson and Seymour use to prove Theorem 5.1. This is not a digression as it leads into our discussion of the proof of Wagner's Conjecture. They use two different approaches depending on whether or not they can find a large clique minor in G. To simplify our discussion of the first technique, we consider an instance (G, X, Y) of k Disjoint Rooted Paths such that G contains not just a large clique minor but, in fact, a large clique. So, let C be a clique in G with 2k + 1 vertices. We will show that some vertex of C is irrelevant. By Menger's theorem we can find (in O(kn) time, using standard techniques) either a set Q of 2k vertex disjoint paths between C and X U Y or a
Brambles And Tree Decompositions
131
set Z of fewer than 2k vertices such that Z separates X U Y from C. In the first case, we can insist that the paths in Q are internally disjoint from C, we simply shorten any path which contains a vertex of C in its interior. Now, the paths in Q along with an appropriately chosen set of edges from C yield the desired k paths between X and Y, see e.g. Figure 14. Thus the (unique) vertex of C which is not the endpoint of any path of Q is irrelevant. (In fact, given this situation, we can simply exit the algorithm and return the desired paths.) X1 1 .....................................................................................
X2 r'
0'.
Y2
Y3
*
.................................................................
Figure 14: c4 is irrelevant
Suppose then that we find a set Z of at most 2k -1 vertices which separates X U Y from the component U of G - Z which contains C - Z. In fact, we can choose an algorithm which returns such a set Z with minimal cardinality as well as a set R of IZI vertex disjoint paths between Z and C. Again, we can take these paths to be internally disjoint from C. We claim that any vertex of C which is not an endpoint of one of these paths is irrelevant. To see this
note that if the desired paths exist in G then their restriction to the graph induced by Z U V(U) is a set of paths with endpoints in Z. But, as in the previous case, any such set of paths will still exist if we delete all of V(C) except those vertices on the paths in R. This implies the claim. A similar approach yields:
Lemma 5.2 ([42]) Let (G, X, Y) be an instance of k Disjoint Rooted Paths and let C be a model of K6k+3 in G. Then there is a vertex of C which is irrelevant. Furthermore, given C, we can find such a vertex in O(kIE(G)J) time.
Showing that there is an irrelevant vertex if G has no clique minor is more complicated. It relies on a characterization of graphs with no large clique
132
B. A. Reed
minor. To give some of the flavour of this characterization, we discuss graphs without Kl as a minor for l E {5, 6}. We recall that no planar graph contains K5 as a minor. Similarly, for every 1 and every surface E in which Ki is not embeddable, no graph embedded in E contains a clique minor of order 1. This, however, is not the whole story. Consider, for example, a graph G obtained from a wall W by adding a vertex adjacent to all of W. This graph contains no K6 minor because W contains no K5 minor. However, for every surface E, if W is high enough then G will not be embeddable on E. Robertson and Seymour [43] managed to prove:
Theorem 5.3 For every 1, if G is a graph with no K1 minor then for every node t of the canonical tree decomposition [T, X] of G, Xt is almost embeddable in some surface on which Kl cannot be drawn.
The definition of almost embedded depends on l and is too complicated
to explain here. We mention only that if for some bounded size set Z of points (the bound depends on 1), G - Z is embeddable in E then G is almost embeddable in E. There are many other allowable extensions of the embedding which we do not mention. We will not be able to say much about the proof of (5.3) since we have not even stated it properly. We remark however that the starting point for the proof is Theorem 4.2. We begin by finding a high wall W which is attached to the canonical tangle T for which Wt (= V(Xt)) is a hitting set. If the rest of the graph is attached to W in a sufficiently non-planar way then G will contain a K` minor. Otherwise, we can extend the embedding of W in the plane to obtain an almost embedding of Xt in a surface in which K` cannot be drawn. Now, (5.3) is important because it allows Robertson and Seymour to apply theorems about graphs actually embedded on surfaces to graphs without a large clique minor. For example, Robertson and Seymour [39] proved (using techniques we will not discuss):
Theorem 5.4 For every surface E and integer k there is an integer h(E, k) such that the following holds. Let (G, X, Y) be an instance of k Disjoint Rooted
Paths such that G is embedded in E. Let W be a wall of height h(E, k) in G. Then, there is a vertex v in W which is irrelevant, and, given W, we can find such a vertex in O(kI E(G) I) time. Using (5.3), Robertson and Seymour [42] extended this to obtain:
Theorem 5.5 For every integer k there is an h(k) such that the following holds. Let (G, X, Y) be an instance of k Disjoint Rooted Paths such that G contains no clique minor of order 6k + 3. Let W be a wall of height h(k) in G. Then, there is a vertex v in W which is irrelevant, and, given W, we can find such a vertex in 0(kI E(G) I) time. More strongly: there is an O(kI E(G) I )time algorithm which given an instance (H, A, B) of k Disjoint Rooted Paths such that H contains a wall of height h(k), finds such a wall as well as either a model of K6k+3 in H or an irrelevant vertex in W.
Brambles And Tree Decompositions
133
Combining (5.5) and (5.2) yields (5.1). This completes our discussion of Robertson and Seymour's algorithm for Disjoint Rooted Paths, we turn now to their proof of Wagner's conjecture. We note first that Kruskal [23] proved 5.6 Wagner's conjecture holds for trees. Robertson and Seymour, using techniques of Nash-Williams [31] extended this to prove:
5.7 Wagner's conjecture holds for graphs of bounded tree width. Now, it is intuitively obvious and not difficult to show:
5.8 For any planar graph H, there is an elementary wall W containing H as a minor.
Proof Idea The most intuitive approach is to draw H in the plane and then to approximate this drawing by paths in a sufficiently fine hexagonal mesh.
From (5.7) and (5.8), we obtain immediately:
Theorem 5.9 ([37]) Wagner's conjecture holds for planar graphs. Proof Consider an infinite sequence G1, G2, ... of planar graphs. We want to show that there exists an i j such that Gi is a minor of G. So, we can assume that for j > 2, G3 does not have G1 as a minor. By (5.8), there exists an h such that every wall of height h contains G1 as a minor. So, we know that for j > 2, G3 does not contain a wall of height h and hence has tree width Now, by (5.7), there are i, j > 2, i j such that Gi is a minor at most of G;. 2534h5.
Robertson and Seymour [40], by bootstrapping with (5.9) and then applying induction, were able to prove:
Theorem 5.10 For every surface E, Wagner's Conjecture holds for graphs embedded on E. With (5.10) in hand, we turn to the proof of Wagner's conjecture for general
graphs. So, consider an infinite sequence G1, G2, ... of graphs. We want to show that there exists an i j such that Gi is a minor of G. Let 1 = IV(Gl)1. For j > 2, let [Ti, Xi] be the canonical tree decomposition of G3. We can assume that for j > 2, Gj does not have G1 as a minor, and hence does not have K1 as a minor. Applying (5.3), we deduce that, for j > 2 and every node t of Ti, Xt is almost embeddable in a surface in which K1 cannot be embedded. This result permitted Robertson and Seymour to extend (5.10) to show that the set of graphs {Xt I j > 2, t E V(TR)} satisfies Wagner's
134
B. A. Reed
conjecture. Thus, the graphs G;, j > 2 can be thought of as trees where the vertices are labelled from a set of graphs satisfying Wagner's Conjecture. Robertson and Seymour, extending Kruskal's theorem once again, managed to show that Wagner's Conjecture must hold for such a set of graphs, and hence
there is an i j such that Gi is a minor of Gj. In the brief account we have given so far, we have glossed over the techniques Robertson and Seymour used to extend Kruskal's theorem (5.6) to obtain first (5.7) and then Wagner's conjecture in general. We close this section by presenting a porous precis of their procedure. One motivation for doing so is simply to deepen the reader's understanding of the approach taken by Robertson and Seymour. Another is that we will need to present some results on tree decompositions which are of interest in their own right. As we saw in Section 2, one property of trees which makes them easy to work with is that they can be rooted. It is this fact which permitted Kruskal to prove (5.6). In particular, he considered rooted minors. We say a rooted tree (Ti, ri) is a rooted minor of a rooted tree (T2, r2) if there is a model of T1 in T2
such that r2 E im(rl). Kruskal actually proved that in any infinite sequence j such that (Ti, ri) is (T1i rl), (T2, r2), ... of rooted trees there exists i a rooted minor of (Tj, rj). Robertson and Seymour extended ideas found in Nash-Williams's proof of this theorem. Crucial to his proof is: 5.11 If (T, r) is a rooted tree, and s is a descendant oft in T, then any rooted tree which is a rooted minor of (Ts, s) is also a rooted minor of (Tt, t). Proof Given a model of a rooted tree (S, p) in (Ts, s), with s E im(p), we add the path of T between s and t to im(P) to obtain a model of (S, p) in (Tt, t) with t E im(p). In order to extend the ideas in Nash-Williams's proof (the details of which we omit) to prove Wagner's conjecture for graphs of bounded tree width, we need to develop a notion of rooted minors for such graphs. We will define this concept in a moment but make some remarks about our definition now. Consider two graphs G1, G2 and corresponding rooted tree decompositions [(T', r1), X1] and [(T2, r2), X2]. In order for [(T1, r1), X1] to be a rooted minor of [(T2, r2), X21 we shall insist that IV(Xr,)l = IV (Xr) I and that there is a model of G1 in G2 such that for each vertex v of Xr, there is a vertex of X,Mz in im(v). Now, in order to mimic Nash-Williams's proof, we need an analogue of (5.11). In order to develop such an analogue, we need to impose more conditions on the tree decomposition and the rooted minors. Specifically: Definition An ordered rooted tree decomposition [(T, r), X] for G consists of a rooted tree, along with a subgraph Xt of G and corresponding ordered set Wt (which is an ordering of V(Xt)) for each node t of the rooted tree such that these subgraphs satisfy the standard axioms for tree decompositions.
Brambles And Tree Decompositions
135
Definition Consider two graphs G1, G2 and corresponding ordered rooted tree decompositions [(T', r1), X1] and [(T2, r2), X2] such that I W,11 I = I W,2I = k.
Let {x1i ... , x,} be the ordering of Wl, and let {yl, ... , yk} be the ordering of W,22. We say that [(T', r1), X1] is an ordered rooted minor of [(T2, r2), X2] if there is a model of G1 in G2 such that for each xi in Wrl, yi is in im(xi). Now, Thomas proved the following beautiful lemma, which as we show below, implies an analogue of (5.11) for ordered rooted minors of certain ordered rooted tree decompositions (Robertson and Seymour proved a weaker version of this lemma which they themselves called a clumsy substitute for it.)
Lemma 5.12 ([49]) Let G be a graph of tree width w. Then there is a tree decomposition [T, X] of G of width w such that the following property holds for every two nodes t1 and t2 of T:
Let P be the unique subpath of T between t1 and t2. If there is
no t on P such that
I Wt I < min (I Wt, Wt2 I) then for every pair of equal sized sets S1 and S2 with Si C Wt; there are IS1I vertex disjoint paths between Si and S2 in G.
Corollary 5.13 Let G be a graph of tree width w. Then there is an ordered rooted tree decomposition [(T, r), X] of G of width w such that ifs is a descendant oft with I W8I = I WtI, and for every node f on the s-t path of T we have IWfI ? I WtI the following holds:
Any ordered rooted tree decomposition which is a rooted ordered minor of [(Ts, s), {Xp I p E V(T8)}] is also a rooted ordered minor of [(Tt,t), {Xp I p E V(Tt)}].
Proof of Corollary 5.13 Let G be a graph of tree width w. Let [T, X] be the tree decomposition G satisfying the condition given in Lemma 5.12. Arbitrarily
choose a root r of T to obtain a rooted tree decomposition [(T, r), W]. Now, we order the vertices of each Wt as follows. We begin at the root, and consider each node before any of its descendants. For a given node a, with IWa) = k say, if there is no ancestor b of a with I Wb I < k then we arbitrarily order the nodes of Wa. Similarily if the first ancestor b of a with (WbI < k encountered on the path from a up to r satisfies I Wb I < k then we order Wa arbitrarily. Otherwise, we let b be the ancestor of a such that I W6I = k and such that no interior node c of the a to b path in T has IWCI = k. We find the set P of k paths between Wa and Wb in G guaranteed to exist by Lemma 5.12. Now,
for each path P in P, if the endpoint of P in Wb is the ith vertex in the ordering of Wb then the endpoint of P in Wa will be the ith vertex in the ordering of Wa. The paths in P ensure that any ordered rooted minor of Wa is an ordered rooted minor of Wb. Corollary 5.13 follows by repeatedly applying this fact to work our way up from s to t.
B. A. Reed
136
Remark The condition in Corollary 5.13 that there is no node f on the s-t path such that Wf has fewer than JW, I elements is obviously necessary, for any such Wf is a cutset in G which may well prevent us from extending our rooted minors. The same holds true for the corresponding condition in Lemma 5.12. Corollary 5.13 allows us to extend Nash-Williams's proof technique to prove
Wagner's Conjecture for graphs of bounded tree width. We turn now to the proof of Wagner's Conjecture for arbitrary graphs. As already noted, we can assume that for some fixed 1 we are considering graphs with no K1 minor. Once again, we want to consider rooted minors, this time using the canonical tree decompositions. Unfortunately, there are two significant complications. The first is that the Wt in such a tree decomposition may not be linked. The second is that the size of the Wt may be arbitrarily large so that, even if we consider one vertex trees, if we choose an infinite sequence of decompositions where every Wt is a different size we will not find a tree decomposition which is a rooted minor of another in the sequence. We deal with these problems by considering rooted minors rooted not at the Wt but in the arcs of the tree, i.e. at W3 flWt for two adjacent nodes s and t
of the tree. This deals quite adequately with the first problem. The fact that the arcs of our tree decompositions correspond to canonical separators allows us easily to prove an analogue of (5.13) for the cutsets corresponding to the arcs. However, these arc cutsets may still be arbitrarily large. To deal with this problem, we simply restrict our attention to the arcs in the canonical tree decompositions which correspond to cutsets with at most f (k) vertices, for a suitably chosen f (k). This yields a new tree decomposition which corresponds to a subset of our original set of laminar separations. We can extend (5.3) to prove that for each node t of this new tree decomposition we still have that Xt is almost embeddable in some surface in which Kl is not embeddable and we are therefore in a position to mimic Nash-Williams's proof. As the above discussion demonstrates, the canonical tree decomposition theorem plays a key role in the proof of Wagner's Conjecture. We expect that it will have many other uses as it seems to be an extremely natural way of decomposing a graph into its highly connected pieces. 6
Packing and covering
Let F be a family of graphs. An F-packing in a graph G is a set of vertex disjoint subgraphs of G, each of which is isomorphic to a member of F. The is the maximum cardinality of an F-packing number of G, denoted F-packing in G. An F-cover is a set X of vertices such that G - X contains no subgraph isomorphic to a member of F. The F-covering number of G, denoted cy(G) is the minimum cardinality of an F-cover for G. Since a cover must contain a vertex from every subgraph of a packing, py(G) < cy(G). Erdos and Posa [15] proved:
Brambles And Tree Decompositions
137
6.1 There exists a constant µ such that for the family C of all cycles, cc (G) < ltpc(G) logpc(G) A family F of graphs is said to have the Erdos-Posa property if there is an integer-valued function f such that cf(G) _< f (pf(G)) for every graph G. Brambles are an extremely effective tool for proving that a given family of graphs has the Erdos-Posa property. For example, Thomassen [50] used them to show that for every integer m, the family of cycles whose length is divisible by m has the Erdos-Posa property. Reed [33] used them to prove that the family of odd cycles has a similar though weaker property (as we shall see the family of odd cycles does not have the Erdos-Posa property). He was thereby able to resolve a conjecture of Erdos [20] concerning odd cycle covers. Reed, Robertson, Seymour, and Thomas [35] used similar techniques to prove that the directed cycles in a directed graph satisfy the Erdos-Posa property. This settled a 25 year old conjecture due to Younger [53]. It seems likely that the theory of brambles will have many other applications to packing and covering. The analogous directed theory will probably also be very useful, once it has been fully developed. We turn now to a more detailed discussion of those applications of brambles to packing and covering mentioned above. In doing so, we will find it convenient to define for every family F of graphs the function f y where for each non-negative integer k, fF(k) is either max{cf(G) I G satisfies pr(G) = k} or oo if no such maximum exists. Then, F has the Erdos-Posa property if and only if f r is integer valued. We note further that for any F, f f(0) = 0. 6.1
Cycles and a useful lemma
In this subsection, we show how brambles (actually, their companions: well-
linked sets) arise when considering the Erdos-Posa property. We then show how they can be used to prove that the Erdos-Posa property holds for the family of all cycles. Suppose that some family .F of connected graphs does not have the Erd6sP6sa property. Then consider the smallest integer k such that fy(k) is infinite.
We know k > 1. Since .F(k) = oo, there is a graph G with pf(G) = k and cf(G) > 4 f f(k - 1). We can apply the following lemma to such a graph.
Lemma 6.2 (The Key Lemma) Consider a family .F of connected graphs, an integer k > 1 such that f ,,(k -1) is finite, and any graph G with pf(G) = k and cf(G) > 4 f f(k - 1). Then every minimum .F-cover H in G is an LIHI/4J linked set.
Proof We need to show that for every set X of at most 1HI/4 vertices of G, some component of G - X contains more than half the vertices of H. So, consider some such set X. Since X is too small to be a hitting set, there must be a component U of G-X containing an element of F. Then, pf(G-U) is at
B. A. Reed
138
most k-1 so there is an F-cover H' for G-U of size at most f,.(k-1) < IHI/4. Now, since H is an F-cover, H fl u is an F-cover for U. Furthermore, since each element of F is connected, X contains a vertex of every subgraph of G which (i) is isomorphic to an element of F, and (ii) intersects both U and G - U. Thus, X U H' U (H fl U) is an F-cover for G. By the minimality of H, it follows that IH fl UI > I HI - I X I - I H'I > IHI/2. This completes the proof of the lemma. To demonstrate the power of this lemma, we apply it to prove: Theorem 6.3 If we let C be the family of all cycles, then fc (k) < 4*
2534(k+1)5
.
Proof Assume the theorem is false and consider the smallest integer k for which the bound claimed for fc(k) fails to hold. Obviously, k > 1. Then there is a graph G with pc(G) = k and cc(G) > 4 * 2534(k+1)5 > 4fc(k - 1). Thus, by the key lemma, G contains a 2534(k+1)5-linked set. Now, Theorems 1.1, 1.2,
and 1.3 imply that G contains a wall of height k + 1. But this wall clearly contains k + 1 vertex disjoint cycles, a contradiction.
Remark Erdos and Posa's original proof gave the much better bound on fc (k) of yk log(k) for a constant p. As we shall see, however, the technique used in our proof can be extended to situations which had previously proven intractable. For example, Robertson and Seymour noted that essentially the same proof shows that for any set S of planar graphs, the family F of graphs which contain some element of S as a minor has the Erdos-Posa property (Erdos and Posa's result is a special case where S consists of one graph, a loop). 6.2
Even cycles
As shown by Thomassen [50], the technique of the last section can be used to prove that the family of even cycles has the Erdos-Posa property. Actually, Thomassen proved that for every integer m, the family of cycles with length 0 mod m satisfy the Erdos-Posa property. In this section, we sketch
Thomassen's proof. We will use the term m-cycles to denote the family of cycles of length 0 mod m.
The result which allowed Thomassen to apply the key lemma to the m-cycles was the following.
6.4 For any h and m (h even), there is a g(h, m) such that if W is a wall of height g(h, m) then W contains a wall W' of height h such that all the paths of W' between nails have length 0 mod m (or equivalently all the paths of W' between two nails and containing no nails in their interior have length 0 mod m).
Brambles And Tree Decompositions
139
Such a W' contains h vertex disjoint cycles. Each of these must be an m-cycle. Thus, we obtain:
6.5 For any h and m (h even), there is a g(h, m) such that if W is a wall of height g(h, m) then W contains h vertex disjoint rn-cycles. Now, we define f (h, m) recursively insisting only that f (0, m) = 0 and for h > 2: f (h, m) > 4f (h - 1, m) + 4 and f (h, m) > 4 * 25349(h+1,"`)5 Mimicking the proof of (6.3), we obtain:
6.6 a graph either has a packing of h vertex disjoint m-cycles or a covering of the m-cycles of size at most f (h, m). Thus the Erdos-Posa property does indeed hold for the m-cycles for every integer m. We remark that Thomassen, in his original proof, considered a tree decomposition of G rather than a k-linked set. We have presented the proof in terms of k-linked sets and brambles for three reasons. The first is to acquaint the reader with these less familiar notions. The second is that k-linked sets seem to generalize to directed graphs more easily than tree decompositions. The third is that, as we shall see in the next section, in some applications it
is important that a minimum cover H is LIHI/4J-linked. Just knowing that G contains a k-linked set for large k may not suffice. 6.3
Odd cycles
In this subsection, we discuss packings and coverings of odd cycles. To begin, we show that the odd cycles do not have the Erdos-Posa property. Our examples are similar to, but slightly different from those given by Lovasz and Schrijver [27]. We remark that Dejter and Neumann-Lara [14] have shown that for any k other than 0, the cycles of length k mod m do not satisfy the Erdos-Posa property, by generalizing these examples. Consider the elementary wall Wh of height h. Wh is bipartite with a unique bipartition, (A, B) say. Both the top and bottom row of Wh contain 2h + 1 vertices. Consider the graph Hh obtained from Wh by first adding a matching between the vertices of degree two in the top and bottom rows apart from the corners so that the 2ith vertex in the top row is joined to the (2h + 2 - 2i)th vertex in the bottom row (see e.g. Figure 15) and then subdividing each of these auxiliary edges, thereby creating h auxiliary vertices. Note that each auxiliary vertex is joined to two vertices of Wh on different sides of the bipartition (A, B). So, the parity of any cycle in Hh depends on the number of auxiliary vertices
it contains. One can use this to show that Hh does not contain two vertex disjoint odd cycles. (One way of doing this is to extend the planar embedding of Wh to an embedding of Hh in the projective plane with all faces even. It follows that any odd cycle of Hh is non-null homotopic. However, any two non-null homotopic cycles in the projective plane intersect.)
B. A. Reed
140
Figure 15: H4
We claim that Hh also contains a set of Lh/2J odd cycles, such that each vertex is in at most 2 of these odd cycles. It follows that Hh has no odd cycle cover with less than Lh/41 vertices and hence the odd cycles do not have the Erdos-Posa property.
To prove our claim, we let vi be the auxiliary vertex joined to the 2ith vertex in the top row, call this ai and the (2h + 2 - 2i)th vertex in the bottom row, call this bi. For 2 < i < h, we let Pi be the path between ai and bi obtained by concatenating (as shown in Figure 16): (i) the subpath of R1 between ai and the last nail before it on R1 which we note is in Ci,
(ii) the subpath of Ci between this nail and some element xi of Ri, (iii) the subpath of Ri between xi and some element yi of Ri n Ch+2_i, (iv) the subpath of Ch+2_i between yi and the first nail of Rh+1 after bi, and
(v) the subpath of Rh+1 between this nail and bi.
Now, Pi + vi induces an odd cycle Ci. Furthermore, a vertex is clearly in at most two of P2, ... , P[h/2j+1 as it is in at most one column and one row. This completes the proof of our claim. In a similar spirit, we define a more general class of counterexamples, the Escher walls.
Definition Let W be a bipartite wall of height h with bipartition (X, Y). Let A = {a1, . . . , ah} be a subset of R1 n X labelled so that ai appears before ai+1
Brambles And Tree Decompositions
141
V4
Figure 16:
on this row. Let B = {b1, ... , bh} be a subset of R2h+1 f1 X labelled so that bi appears after bi+1 on this row. Let A be such that for 1 < i < h there is some internal vertex of the subpath of R1 between ai and ai+1 which is a nail. Similarly, let B be such that for 1 < i < h there is some internal vertex of the subpath of R2h+1 between bi and bi+1 which is a nail. Let P1, ..., Ph be vertex disjoint paths such that Pi has endpoints ai and bi, is internally vertex disjoint from W, and has an odd number of edges. Then W U P1 U ... U Ph forms an Escher wall of height h. The same proof techniques show that an Escher wall of height h contains no 2 vertex disjoint cycles but does contain Lh/2J odd cycles using each vertex at most twice and hence no odd cycle cover with fewer than Lh/4J vertices. Thus, the Escher walls provide a counterexample to the statement: the Erdos-Posa property holds for the odd cycles.
The result we discuss in this section shows that, in a certain sense, they are the only counterexamples. To wit,
Theorem 6.7 For every h and k there is an f (h, k) such that every graph contains either k vertex disjoint odd cycles, an Escher wall of height h, or an odd cycle cover with at most f (h, k) vertices.
This theorem has two interesting consequences. First, it implies the following result relating half-integral packings and coverings of odd cycles.
Definition A half-integral packing of odd cycles in G consists of a set S of odd cycles such that each vertex of G is in at most two elements of S. The size of the packing is IS112.
B. A. Reed
142
Corollary 6.8 For every k there is a g(k) such that every graph contains either a half-integral packing of odd cycles of size k or an odd cycle cover with at most g(k) vertices.
Proof Simply set g(k) = f (4k, k). Since an Escher wall of height 4k contains a half-integral packing of size k, the result follows. Theorem 6.7 also implies the following conjecture of Erdos.
Definition A graph is k-quasi bipartite if every subgraph H of G contains a stable set with at least IV(H)1/2 - k vertices. (Note that a bipartite graph is 0-quasi bipartite, we simply pick the larger side of the bipartition of H. Since an odd cycle with 21 + 1 vertices contains no stable set of size 1 + 1 it follows that a graph G is 0-quasi bipartite if and only if it is bipartite.)
Conjecture 6.9 (Erdos [20]) For all k there is an 1(k) such that every k-quasi bipartite graph contains an odd cycle cover of size at most 1(k). Remark Erdos's original conjecture concerned k-near bipartite graphs in which only the even H needed to contain large stable sets. Since every k-near bipartite graph is (k + 1)-quasi bipartite, the conjectures are equivalent.
Proof of Conjecture 6.9 We need the following result which is proved using brute force. Observation 6.10 For every k there is an h(k) such that an Escher wall of height h(k) is not k-quasi bipartite. We also need the trivial fact that a set of 2k + 1 vertex disjoint odd cycles is not k-quasi bipartite. Now, Erdos's Conjecture follows from Theorem 6.7 with 1(k) = f (h(k), 2k + 1).
The proof of (6.7) uses the fact that under certain conditions, a minimal IH1/4]-linked set. However, hitting set H for the odd cycles in a graph G is a [ we actually consider a bipartite graph obtained from G and H as follows.
Definition Let H = {x', ... , xI HI } be an odd cycle cover in a graph G. Let (A, B) be a bipartition of G - H. For each vertex x of H, we define two new vertices XA and XB. We let H' = U{XA, xB I x E H}. We define a new graph G' = G'(H, (A, B)) with vertex set V - H U H' and edge set
E(G - H) U {xAb I b E B, xb E E(G)}
U{xBaI aEA, xaEE(G)}U{xAxBIxixj EE(G), i<j}.
Brambles And Tree Decompositions
143
Now, we prove (6.7) for each h in turn. Thus, we restrict our attention to the class of graphs with no Escher wall of height h and prove that the Erd6sP6sa property holds for this class of graphs. Hence, we can apply the Key Lemma to prove that a minimum hitting set H for an appropriate counterLIHI/4J-linked. It turns out that similar techniques allow us to example G is prove that H' is LIH'1/4J-linked in G'. Now, (4.2) implies the existence of a large wall W in G' attached to the tangle defined by this LIH'I/4J-linked set. Thus, for every small cutset X the component of G - X completely containing a row of W also contains at least half the vertices of H'. It is this fact which allows us to find a set of paths between some elements of H' and the nails on the perimeter of some subwall W' of W such that these paths together with W' either yield an Escher wall of height h in G (a contradiction), or contain the desired k vertex disjoint odd cycles. The proof is too complicated to discuss further, we simply remark that the following easy observation links it to the Disjoint Rooted Paths problem and allows us to apply techniques Robertson and Seymour developed for that problem.
Observation 6.11 Let x1, ... , xL be 1 elements of H. let P1, ..., Pa be vertex disjoint paths of G' such that Pi has endpoints xA, xB and is internally disjoint from H. Then, these 1 paths correspond to 1 vertex disjoint odd cycles in G. 6.4
Directed cycles
In 1966, Gallai [19] conjectured that there is an integer n such that any directed graph D without two vertex disjoint directed cycles contains a set X of at most n vertices such that D - X is acyclic. In 1973, Younger [53] conjectured that for every k there is an f (k) such that every directed graph contains either k vertex disjoint directed cycles or a set X of at most f (k) vertices such that D - X is acyclic. In 1991, McCuaig [29] proved Gallai's conjecture, i.e. Younger's conjecture for k = 2, with n = 3. Until recently, Younger's conjecture has remained open for all other values of k except the trivial cases k = 0 and k = 1. In 1995, Reed, Robertson, Seymour and Thomas [35] proved Younger's conjecture. To do so, they considered directed analogues of well-linked sets and grids. Definition A set S of vertices of a digraph D is well-linked if for every pair X and Y of equal-sized disjoint subsets of S there exist IXI vertex disjoint paths from X to Y in D - (S - X - Y). (Note that by symmetry we also have paths from Y to X. Note also that the definition is the same if we do not insist that X and Y are disjoint.)
Definition A directed wall of height h (see Figure 17) consists of two sets of h + 1 disjoint paths R1, ... , Rh+1 and C1, ... , Ch+1 such that each Rz and C; intersect in a non-empty path and furthermore:
B. A. Reed
144
`2
fCi
ci4
Figure 17: A directed wall of height 4
(i) Each Ri intersects the C; in increasing index order, (ii) Each C; intersects the R;, in increasing index order. For convenience, we also define fd,(k) to be the maximum cardinality of a minimum directed cycle cover over all graphs G without k + 1 vertex disjoint cycles, or oo if this maximum is undefined.
Although one cannot prove that a minimum hitting set of the directed cycles in a directed graph is well-linked, the following fact is true and can be proved along the same lines as the Key Lemma:
Fact 6.12 Let D be a digraph without k+1 vertex disjoint directed cycles. Let H be a minimum hitting set for the directed cycles of D. Let X and Y be any two disjoint equal-sized subsets of H, each containing at least fdc(k-1) vertices.
Then there are IXI vertex disjoint paths from X to Y in D - (H - X - Y). Thus, we see that a minimum hitting set for the directed cycles in a directed graph is well-linked modulo its small subsets. Now, Reed, Robertson, Seymour
and Thomas were able to show that for any h and 1, if a digraph contains a sufficiently large set which is well-linked modulo its subsets of size at most 1 then it contains either (i) h vertex disjoint directed cycles, or
(ii) a directed wall W of height h as well as h vertex disjoint paths Ql, ... , Qh such that Qi is a path from the last vertex of Ci to the first vertex of Ci (which is not neccesarily internally vertex disjoint from W).
They also proved, using straightforward induction arguments, that for every k there is an f (k) such that if a graph contains a directed wall of height f (k) and corresponding paths as in (ii) then it contains k vertex disjoint cycles.
Combining these three results yields a proof of Younger's conjecture.
The proof suggests that the directed analogue of tree width may have many applications. Unfortunately, at the moment even the correct definition of directed tree width seems hard to determine.
Brambles And Tree Decompositions
145
However, the following conjecture, a directed analogue of (1.2), seems to be of interest and its proof may well lead to a better understanding of global directed connectivity.
Figure 18: A directed cylinder of height 3
Definition A directed cylinder of height h (see Figure 18) consists of a set of 2h disjoint paths R1, ... , R2h and a set of h disjoint directed cycles C1, ... , Ch such that each Rz and C3 intersect in a non-empty path and furthermore: (i) If i < h then R$ intersects the Cj in increasing index order, (i) If i > h then Rt intersects the C3 in decreasing index order,
(ii) For each C3 there is an arc ej of C3 such that the path Pj = C3 - ej intersects the Rz in increasing index order.
Conjecture 6.13 For every h there is an f (h) such that if a directed graph contains a well-linked set of order f (h) then it contains a directed cylinder of height h. We present this conjecture at the end of our discussion of brambles and tree decompositions. However, we hope that it is the beginning of a rich theory of directed brambles and directed tree decompositions.
7
Building a wall
In this section, we prove Theorem 1.2 which states that if a graph G has no wall of height g then it has no bramble of order greater than 253495 - 1. The
B. A. Reed
146
proof we present is long and fairly technical. We foist the details on the reader for three reasons. The first is that they require an application of the Closest Cutter Lemma discussed in Section 3 and hence illustrate the importance of the laminarity of cutters. The second is that, as we saw in the last two sections,
the theorem has many important applications. The third is that our proof although similar to that given in [46], avoids the use of a result from matroid theory, and uses tangles rather than preferences. We hope that this makes it more accessible to the reader. As mentioned in Section 4, the main idea of the proof is to start with a large 7--linked set for some huge order tangle T and then grow a set P of paths
from this set "towards" the tangle. These paths shall form the rows of our wall. We shall show that we can also finds a set C of disjoint paths between the elements of P which we will use to form the columns. We turn now to the details. (4,2)
(4,3)
(3,1)
(3, 2)
(3,3)
(2,
1)
(2,2)
(2,3)
(2,4)
(2,5)
(17
1)
(1,2)
(1, 3)
(1,k4)
(1,5)
(4,4)
V V/ (0,2)
(0,3)
(0,4)
(0,5)
Figure 19: L4,5
We begin with a definition. We denote by Lh,k the graph with vertex set {(x) y) 10 _< x < h, 1 < y < k} where (0, y) is adjacent to (x, y) for 1 < y < k,
1<x 2. Let T be a tangle of order 025nh in a graph G, and let Z be a T-linked set of order h in G. Then, there are subgraphs H and G' of G such that setting Z' = V (H) flV (G') we have:
(i) (H, Z, Z') is an (h, n)-expander, and (ii) there is a tangle ' in G' of order 0 and conformal with T such that Z' is a T'-linked set.
Lemma 7.3 Let g, h and 0 be integers with g > 2, h > 3 and 0 > 20gh. Let n = (2g + 7) (2h - 5) + 2 (note that 0 > n). Let T be a tangle of order 20254g(h-1)+n in a graph G containing no wall of height g. Let Z be a T-linked set of order n in G. Then, there are subgraphs H and G' of G such that setting Z'= V(H) fl V (G') we have:
(i) there is a tree Tin H-Z' such that (H, Z, Z', T) is an (n, h)-concentrator, and
(ii) there is a tangle T' in G' of order 0 and conformal with T such that Z' is a T'-linked set.
Now, as we have remarked, if (H, X1i X2) is an (h, n)-expander and if (F, X2, X3, T) is an (n, h)-concentrator such that V (H) fl V (F) = X2, then (H U F, X1, X3, T) is an h-step. Thus, Lemmas 7.2 and 7.3 taken together imply
Lemma 7.4 Let 0, g and h be integers with g > 2 and h > 3 and 0 > 20gh. Let n = (2g + 7)(2h - 5) + 2 (note 0 > n). Let T be a tangle of order a graph G containing no wall of height g. Let Z be a Tlinked set of order h in G. Then, there are subgraphs H and G' of G such that 2025nh+4g(h-1)+n in
setting Z' = V (H) fl V (G') we have:
(i) there is a tree T in H - Z' such that (H, Z, Z', T) is an h-step, and (ii) there is a tangle T' in G' of order 0 and conformal with T such that Z' is a T'-linked set.
Now, by (2.13) we know that every proper subset of a minimum order hitting set for T is a 7--linked set. Thus, if G contains a tangle T of order at least 0 then it contains a T-linked set of order at least 0. This fact, combined with k applications of Lemma 7.4 and our earlier remarks about finding an Lh,k minor in a sequence of k h-steps yields
Brambles And Tree Decompositions
149
Theorem 7.5 Let g, h, and k be integers with g > 2 and h > 3. Let n = (2g + 7) (2h - 5) + 2. If T is a tangle of order at least in a graph G which contains no wall of height g then G has an Lh,k minor.
20gh25(nh+4g(h-1)+n+1)k
Since Lg+1,92 contains a wall of height h we obtain:
Corollary 7.6 If G has a tangle of order at least 253495-1 then it contains a wall of height h.
Proof For g > 2 this follows from (7.5). If a graph has no wall of height 1, then it contains no cycle of length 6 or greater. This implies that each of its blocks is a subdivision of a multigraph with at most 5 vertices. It is easy to construct a tree decomposition of any such graph of width at most 5. The corollary follows.
Since every bramble of order 3k defines a tangle of order k, this yields:
Corollary 7.7 If G has a bramble of order at least 253495 then it contains a wall of height h.
It remains only to prove Lemmas 7.2 and 7.3. The following two technical
results whose proofs appear at the end of this section, are the keys to these two lemmas.
Lemma 7.8 Let 0 > 2 be an integer. Let T be a tangle of order at least 240 + 7 in a graph G. Let Is, t} be a T-linked set in G. Then there is a path P between s and t such that there is a tangle in G - V(P) of order 0, conformal with T. Lemma 7.9 Let C be a connected subgraph of a graph G, let h be an integer, and let T be a tangle of order greater than 2h in G - C which we call To when
thinking of it as a tangle in G. Let Z be a To-linked set of order h which intersects C. Then G contains two subgraphs G' and H such that setting Z' = V(H) n V (G') we have: (i) IZ'I = h and there are h vertex disjoint paths in H between Z and Z',
(ii) there is a tangle T' in G' of order ord(T') - h and conformal with T such that Z' is a T-linked set, and
(iii) G'nC=O. To illustrate the importance of Lemmas 7.8 and 7.9 we use them to deduce the following two lemmas, from which Lemma 7.2 follows almost immediately.
B. A. Reed
150
Lemma 4.1 Let h and 0 be integers with 0 > h > 2. Let T be a tangle in a graph G of order at least 24(h + 0) + 7, and let Z be a T-linked set of order h in G. Let s and t be two vertices of Z. Then, there are subgraphs H and G' of G such that setting Z' = V(H) n V (G') we have:
(i) JZ'J = h and there are h vertex-disjoint paths P1, ..., Ph in H from Z
to Z' and an s to t path Q in H - Z' such that for each i, Q n Pi is a (possibly empty) path, and (ii) there is a tangle T' in G' of order 0 and conformal with T such that Z' is a T'-linked set.
Proof By Lemma 7.8, there is a path C between s and t and a tangle Ti in G - V(C) of order h + 0 conformal with T. So, we have that (in G) Ti is indistinguishable from T. We claim that (in G) Z is a Ti-linked set. Otherwise, there is a set X with fewer than IZI vertices separating Z from some hitting set H for T1. Since ord(Ti) > I Z I > IX1, there is some component U of G - X
containing an element T of Ti. Since T intersects H, U n Z = 0. Since T is indistinguishable from Ti, we know there is an element of T contained in U. It follows that every element of T intersects U + X, and hence there is a hitting
set H' for T in U + X. But then X is a cutset separating Z from H'. This contradicts the fact that Z is T-linked, proving the claim. Since ord(Ti) > 21ZI, we may apply Lemma 7.9. We deduce that there are subgraphs H and G' of G such that setting Z' = V (H) n V (G'), we have: (i) JZ'J = h and there are h vertex disjoint paths P1,
...,
Ph in H between
Z and Z', (ii) there is a tangle T' in G' of order ord(T') - h = 0 and conformal with T such that Z' is a T-linked set, and
(iii) G' n C = 0.
Now, we let Q be a path in H - Z' between s and t which minimizes JE(Q) -Ui E(Pi) 1. Such a choice is possible because Q = C is one possibility. Clearly, Q intersects each Pi in a path, as required.
Lemma 7.10 Let 0 and n be integers with 0 > n > 2. Let T be a tangle of order at least 025" in a graph G, and let Z be a T-linked set in G of order n. Then, there are two subgraphs H and G* of G such that setting Z* = V(H) n V (G*), we have:
(i) IZ*I = n, and there are n vertex disjoint paths from Z to Z' in H,
(ii) there is a tangle T* in G* of order 0 and conformal with T such that Z* is a T*-linked set, and (iii) Z is in one component of H - Z*.
Brambles And Tree Decompositions
151
Proof We prove that for each r between 0 and n - 1 there are subgraphs Hr and G'r of G such that setting Zr = V(H,) fl V(G,) we have that: Hr-1 C H,., IZrI = n and there are n vertex-disjoint paths from Z to Zr, there is a tangle T in G'r of order 925" and conformal with T such that Zr is a T-linked set, and Z is in the union of n - r components of Hr - Zr. The lemma follows as we simply set Z* = Z"_1, G* = G;i_1i and T* = T"_1.
We begin by setting Zo = Z, G'0 = G, To = T, and Ho to be the graph induced by Z. For r > 1, having obtained Zr_i, Hr_1, G' _1 and T,-,, we obtain Zr, Hr, Gr and T by applying Lemma 4.1 to Zr_l, Hr_1, G;._1 and T,-1 and a pair of vertices s and t of Zr_1 which are in different components of Hr-1 (unless all of Zr_1 is in the same component of Hr_1 in which case s and t are chosen arbitrarily from Hr_1). This yields two subgraphs H and G' in Gr, and
a tangle T' in G'. We set Hr = H U Hr-1, Gr = G', Zr = V(H) fl V (G), and
T =T'. We now deduce Lemma 7.2.
Proof of Lemma 7.2 Let h, 0, n, G, T and Z be as in the statement of the lemma. We know that Z is contained in a T-linked set Zo of order n. We let To = T and Go = G. We shall find a sequence of sets Z1, ..., Zh, two sequences of graphs H1, ..., Hh and G1, ..., Gh and a sequence of tangles T1, ... , Th such that for i E {1,. .. , h}: (i) V (Hi) n V (Gi) = Zi,
(ii) Hi C Gi_1, (iii) IZiI = n and there are n vertex disjoint paths from Zo to Zi in Uj=1 Hj C
G - Gi, (iv) Zi_1 is contained in one component of Hi - Zi, and (v) Ti is a tangle of order at least 025(h-i)" in Gi conformal with T and Zi is a Ti-linked set. We find the desired Zi, Hi, Gi, Ti using h applications of Lemma 7.10. We claim that setting Z' = Zh, G' = Gh , H = Uh 1 Hi , T' = T" proves the lemma. We need only show that for every set Y in Z' with FYI = h there are h vertex disjoint paths from Z to Y in H. If the paths did not exist then there would be a set X of at most h - 1 vertices which separates Z from Y. Now, we know there are n vertex disjoint paths from Zo to Z' in H. Clearly, there
is one such path P which is disjoint from X and has an endpoint in Z, and another P' which is disjoint from X and has an endpoint in Y. Furthermore, there is some H1 with H1 fl X = 0. However, both P and P' contain a vertex of Z1_1 and Z1_1 lies in one component of H1, and hence of H - X. This contradicts our assumption that X separates Z from Y in H.
B. A. Reed
152
The proof of Lemma 7.3 requires a bit more work. In particular, we will need the following lemma.
Lemma 7.11 Let 0, g, and n be integers with 0 > n > (2g + 5). Let T be a tangle of order at least 025 4g in a graph G which contains no wall of height g. Let Z be a T-linked set of order n in G and let {v1, ... , v2g+5} be some ordered
set of vertices of Z. Then there are graphs F and G* in G such that setting Z* = V(F) f1 V (G*), we have: (i) I Z*I = n and there is a set P of n vertex disjoint paths from Z to Z*
in F, (ii) there is some 1 E {3, ... , 2g + 3} such that letting P1 be the element of P which has v1 as an endpoint, there is some other element P2 of P which has neither v1_1 nor v1+1 as an endpoint and a path Q of F - l f {V (P) I P E P - P1 - P2 } which has an endpoint on P1 and an endpoint on P2,
(iii) there is a tangle T* in G* of order at least B and conformal with T for which Z* is a T*-linked set.
Proof We define Zo = {z°, ... ,
Z, To = T, and Go = G. Form E
{1, ... , 2g + 5}, we define k,,,, so that zkm =
We apply Lemma 4.1, 4g - 1 times to prove the existence of a sequence of graphs G1 D . . . D G4g_1i a sequence of tangles T1, ... , Tg_,, and a sequence of sets of vertices Z1,... , Z4g_1 with Zi = {zi, ... , z;,}, such that for 1 < i < 4g - 1
(i) Zi_1 C_ Gi_1 - Gi and there is a set Pi = {Pi,1, ... , P1,n} of n vertex disjoint paths in Gi_1 - (Gi - Zi) from Zi_1 to Zi such that Pi,j links to (ii) Ti is a tangle of order at least B254g-i in Gi conformal with T and Zi is a Ti-linked set of Gi, (iii) there is a path Qi of Gi_1- Gi between zk9+3 and zip 1 whose intersection with every element of Pi is a (possibly empty) path.
For each j in 1, ... , n we let Rj be the path from zio to z11-1 obtained by concatenating {Pi,j 1 < i < 4g - 1}. Now, if for any Qi there is a 1
subpath Q of Qi which has its endpoints on Rk, and R, with 1 E {3, ... , 2g + 3} and r 0 {k1_1, k1, k1+1} then setting Z* = Z4g-1, T* = Tg_1, G* = G4g_11
F = G - (G4g_1 - Z4g_1), and P = {Rj I 1 < j < n} shows that the lemma holds. So, we can assume that each Qi either intersects Rk9+a, Rk9+4, . . . , Rk2g+3 in the given order before intersecting any other Rj, or intersects Rk9+3, Rk9+2
... , Rk3 in the given order before intersecting any other Rj. Now, either there are 2g values of i such that Qi satisfies the first condition or there are 2g values
Brambles And Tree Decompositions
153
of i such that Qi satisfies the second condition. In the first case, the subgraph consisting of these Qi along with Rk9}3 7 ...I Rk,g13 clearly contains a wall of height g. Similarily, in the second case, there is a wall of height g in G. But G contains no wall of height g so the lemma holds.
Proof of Lemma 7.3 Let g, h, n, 0, G, T and Z be as in the statement of Lemma 7.3. Set Zo = Z, Go = G, and To = T. By Lemma 7.10, we can find an Hl, G1, Z1, Ti, and Ti such that V (Hl) n V(Gi) = Z1, Tl is a tree in Hl - Z1, Z1 I = n, there are n vertex-disjoint paths P1, ..., Pn in H1 from Zo to Z1 such that T1 intersects each Pi and Ti is a tangle in G1 of order 202549(h-1) conformal with T such that Z1 is a T1 linked set. By choosing T1 so as to minimize E(T) - U 1 E(Pi) we can ensure that T1 intersects each Pi in a path. Furthermore we can also ensure that each leaf 1 of T1 is T fl Pi for some i.
We prove by induction on r that for each r between 1 and h - 1 we can find subgraphs Hr and Gr of G such that setting Zi = V (Hi) fl V (Gi) we have:
(1) Hr-1 C Hr, (2) JZ J = n and there is a set Pr of n vertex disjoint paths Pr,l, ... , of Hr between Z and Zr,
Pr,n
(3) there is a tree Tr in Hr-Zr with at least r + 1 leaves intersecting each Pr,i in a non-empty path such that each leaf of Tr is the intersection of Tr with some Pr,j, and
(4) there is a tangle T of order such that Zr is a T-linked set.
202549(h-r-1) in Gr and conformal with
T
This implies Lemma 7.3 as can be seen by letting Vi, ... , Vh be h leaves of Th_1i letting zi be the vertex of Zh_1 which is on the path in Ph_1 which contains vi, letting T = Th_1 - {v1, ... , Vh}, letting Z' = {z1, ... , zh}, letting
G' = Gh_1- (Zr_1- Z'), letting H = Hh_1, and letting T = T-1/(Zh-1- Z'). Now, we see that the result holds for r = 1. To prove the existence of Hr and Gr satisfying (1)-(4) given Hr-1 and Gr_1i we will apply Lemma 7.11 to Zr_1 in Gr_l. To this end, consider the tree Tr_l. We can assume that Tr_1 has no more than r leaves as otherwise, we simply set Tr = Tr_i, Gr = Gr_1,
Hr = Hr-1, T = T_1, and Pr,j = Pr_i,j. So, Tr_1 has exactly r leaves and can therefore be partitioned into at most 2r - 3 paths, all the internal vertices of which have degree 2 in Tr_1. Now, one of these paths R intersects at least n/(2r - 3) > 2g + 7 of the paths in Pr_i, and hence there are at least 2g + 5 paths of Pr_1 whose intersection with Tr_1 is contained in the interior of R.
Let L1, ..., L29+5 be the first 2g + 5 such paths, enumerated in the order that they appear along R. Let wi be the endpoint of Li in Zn_l. We apply Lemma 7.11 to Zr_1, T,-,, Gr_1, and {w1, ... , w29+5} to obtain F, G*, T*, Z*, w`, P and Q. We set Zr = Z*, Gr = G*, T = T*, Hr = Hr-1 U F, we
B. A. Reed
154
concatenate Pr_1,j with an element of P to obtain Pr,j, and we use T,._1 and Q to create Tr. To this end, let R1 and R2 be the paths of Pr containing the endpoints of Q with L` C R1. Note that there is a path Q' in Q U Rl U R2 between a vertex of L` fl Tr_1 and R2 fl Tr_1 with no internal vertex in Tr_1 which intersects both R1 and R2 in a path. Let C be the unique cycle of Tr_1 U Q'. We know that C contains either a subpath of R between L`_2 and L`_1 or a subpath of R between L1+1 and Li+2. By symmetry, we may assume the second possibility occurs. Let w be the endpoint of Q' not in LI. Assume first that w is not in LI+2. Then letting Q* be the maximal subpath of R with one endpoint x in L1+1 and the other y in L1+2 we see that letting Tr be the tree T,-,+Q'- (Q* -x-y) yields the desired result (as x and y are leaves of Tr which are not leaves of Tr_1 and every leaf of T,_1 with the possible exception of w is a leaf of Tr). Otherwise, w is in Rfl L1+2 and hence is not a leaf of Tr_1.
So, letting Q* be a maximal path of R with one endpoint w and the other x in L1+1 and setting Tr = Tr_1 + Q - (Q* - w - x) yields the desired result.
Proof of Lemma 7.9 The proof of Lemma 7.9 relies on the following easy corollary of the Closest Cutter Lemma.
Lemma 7.12 (The Relinking Lemma) Let T be a tangle of order at least h + 2 in a graph G. Let Z be a T-linked set in G with h vertices. Let z be a vertex of Z. Then there is some neighbour y of z such that Z - z + y is a Tlz-linked set in G - z.
Proof Since Z is T-linked, Z - z is T/z-linked in G - z. So, by the Closest Cutter Lemma, there is a set X * of h - 1 vertices of G - z such that X* separates Z-z from T/z and such that any set Y of h-1 vertices of G - z separating Z - z from T/z is disjoint from fT/,z(X*). Note that fr/z(X*) = fT(X*). Thus, because Z is T-linked, there is a neighbour y of z in fT/,,(X*). By our choice of X*, Z - z + y is T/z-linked. Now, repeated applications of the Relinking Lemma yield:
Lemma 7.13 (The Linked Set Moving Lemma) Let T be a tangle of order at least 2h+ 1 in a graph G. Let Z = {z1, ... , zh} be a T-linked set in G with h vertices. Then there is a set Y = {y1, y2, ... , AI of vertices of G - Z such that yi is a neighbour of zi and Y is a T/Z-linked
set in G - Z. With these lemmas in hand we are ready to prove Lemma 7.9. So, let h, G, T, Z, and C be as in the statement of the lemma. By induction, we can assume the statement holds for any minor of G. By the Closest Cutter Lemma, there is a T-linked set X* such that X * is a set of h vertices separating Z from T, any set of h vertices separating Z from T is disjoint from fy(X*), and there are h vertex disjoint paths between Z and X*. We consider two cases.
Brambles And Tree Decompositions
155
Case 1: C fl fy(X*) = 0. In this case, by the Linked Set Moving Lemma, there exists a T/X* linked set Y of h vertices in G - X* such that there is a matching between Y and X*. Obviously, Y C_ fy(X*) C G - C. Furthermore, the paths from Z to X* are disjoint from fT(X*) and T/X* is contained in fT(X*). So, setting G' = fT(X*), H equal to the subgraph of G induced
by G - (fT(X*) - Y), Z' = Y, and T' = T/X* we see that Lemma 7.9 holds.
Case 2: C fl fy(X*) # 0. In this case, since C is connected and C fl Z # 0, there is an edge e of C with at least one endpoint in fy(X*). Let G* be the graph obtained by contracting e. Let Z* consist of those vertices of Z in G* and the vertex of G* not in G if one of the endpoints of e was in Z. Clearly T is a tangle in G* - C = G - C. We claim that Z* is a T-linked set in G*. Otherwise,
there is a set Y of at most h - 1 vertices of G* separating Z* from T. Now, Y corresponds to a set of vertices of G which separate Z from T. This set is either Y or is a set of h vertices containing both endpoints of e. In the first case, we contradict the fact that Z is T-linked. In the second
case, we contradict the choice of X*. So, by the induction hypothesis, we can find subgraphs H and G' of G* as described in Lemma 7.9. By "uncontracting" the edge e, we convert these to subgraphs which show that Lemma 7.9 holds for G.
Proof of Lemma 7.8 We begin with the following lemma.
Lemma 7.14 Let h > 2 be an integer, let G be a graph, let Z be a well-linked set of 24h + 6 vertices of G, let (S, T) be a partition of Z with
BSI = ITI = 12h + 3. Then there is a path P with one endpoint in S, the other in T, and otherwise disjoint from Z such that for any set Y of fewer than h + 2 vertices of G - P, there is a component of G - P - Y which contains at least 16h + 3 vertices of Z.
Proof Let T be the tangle consisting of those connected subgraphs of G containing more than two thirds of the vertices of Z. Because Z is welllinked, T must have order at least 8h + 2. To begin, we show: (1) If there is a path P from S to T and a set X of at most 7h vertices such that X separates P from T, then P shows that the lemma is true.
Proof of (1): To see this note that for any set Y of less than h + 2 vertices of G none of which are on P, fT(X UY) C fy(X) and hence is disjoint from P. So, fT(X U Y) is contained in a component of G - P - Y which therefore contains at least 16h + 3 vertices of Z. 11
B. A. Reed
156
Now, let P be a set of 12h + 3 vertex-disjoint paths from S to T in G. Then, every element of P is internally vertex-disjoint from Z and we shall prove that one of them satisfies the theorem. We say a path P1 in P splits a pair (P2, P3) of paths in P, if there do not exist h + 2 vertex disjoint paths from P2 to P3 in G - P1. We may assume that: (2) For any three distinct paths P1, P2, and P3 of P, if P1 splits (P2, P3) then P2 does not split (P1, P3).
Proof of (2): If there do not exist 7h vertex disjoint paths from P3 to P1 then (1) implies that one of P1 or P3 shows that the lemma is true. So, we can assume such paths exist. Each of these paths either hits a vertex of P2 first or a vertex of P1 first. The result follows. By (2), there are at most 123+3) pairs {P1, (P2, P3)} such that P1, P2,
and P3 are distinct paths of P and P1 splits (P2, P3). It follows that there is some path P of P which splits at most (12h + 2)(12h + 1)/6 unordered pairs of paths of P - P. We claim that P shows the lemma is true. For, if not then there is a set Y of less than h + 2 vertices of G - P such that every component of G - Y - P contains fewer than 16h + 3 vertices of Z. It follows that there is a separation (A, B) of G - P with V (A n B) = Y such that A and B both contain at least 8h + 2 vertices of Z - P. Since JYJ < h + 2, there are at least llh + 1 paths of P - P which are either wholly within A or wholly within B. Furthermore, there
must be at least 7h/2 paths of P - P wholly within A and 7h/2 paths of P - P wholly within B. Also, P splits any pair of paths, one of which is wholly within A and the other of which is wholly within B. It follows that P splits at least 7h(15h + 2)/4 paths, a contradiction.
We turn now to the proof of Lemma 7.8. So, we let 0, G, T, is, t} be as in the statement of the lemma. We choose a set Z of at most 240 + 6 vertices of G, a partition (S, T) of Z, and two disjoint connected subgraphs J and K of G - f7-(Z) such that (1) Z is T-linked, (2) ISI, ITI < 120 + 3,
(3) s E V(J), t E V(K), J n z is a vertex of S, K n Z is a vertex of T, and (4) every vertex of S - J is adjacent to a vertex of J, every vertex of T - K is adjacent to a vertex of K. Furthermore, subject to (1)-(4) we choose Z so as to minimize f7-(Z). Such
a choice is possible because setting Z = Is, t}, J = S = s, K = T = t, satisfies (1)-(4). We claim that:
Brambles And Tree Decompositions
157
(5) there is no set Y of IZI vertices which separates Z from T.
Proof of (5): If such a Y exists then by the Closest Cutter Lemma, we can choose Y so that there are r = JZJ vertex disjoint paths P1, ..., P, from Z to Y. W e enumerate Z as v1, ... , v, and Y as w1, ... , w, so that Pi has endpoints vi and wi. Let J' be the union of J and
U Pi - wi + {wi I vi is the unique vertex of J fl s} , vi ES
then J' is connected by (4). Let K' be the union of K and
U Pi - wi + {wi I vi is the unique vertex of K n TI, vi ET
then K' is also connected by (4). Let S' = {wi vi E S} and let T' = {wi I Vi E T} . Then, Y, S', T', J', K' contradicts the minimality of ffr(Z) I
From (5), we obtain easily (6) ISI = ITI = 120 + 3.
Proof of (6): Suppose S contains fewer than 120+3 vertices. Then let v be the unique vertex in J fl S. Since Z is T-linked, v must have a neighbour w
in fy(Z). Furthermore, by (5), Z+v must be T-linked. But now, Z+v, S + v, T, J, K contradicts the minimality of fy(Z). So ISI = 120 + 3. Similarily, we obtain ITI = 120 + 3, proving (6).
We are now in a position to define a tangle in fy(Z) U Z, using: (7) Z is a well-linked set in the graph G' induced by Z U f7-(Z).
Proof of (7): Suppose not, then there is a separation (A, B) of G' such that setting Y = V (A fl B), we have IY U (A fl Z) 1, IY U (B fl Z)
240 + 5.
Let Y'= (AnZ) UY. Let Y"= (BnZ)UY. Note that, by (1), fT(Y'.) ¢ fT(Z) and fy(Y") ¢ fT(Z). However, clearly if fy(Y') intersects A then it is contained in A - Z C fT(Z), a contradiction. So, fT(Y') n A = 0. Similarly, fT(Y") n B = 0. So, fT(Y') n fT(Y") n fT(Z) = 0 and further any edge with an endpoint in fT(Y') fl fT(Y") is disjoint from fT(Z). But this contradicts the fact that T is a tangle (consider the tangle elements in fT(Y'), fy(Y"), and
fr(Z)) Now, from (7), (6), and Lemma 7.14 we obtain:
B. A. Reed
158
(8) There is a path P' of G' from S to T, with no internal vertex in Z such that for every set Y of less than 0 + 2 vertices of G' some component of G' - Y - P' contains more than 160 + 2 vertices of Z - P.
Now, let P be a path of G from s to t whose vertex set is contained in J U K U P' (such a path exists by (4)). By (3), 1Z - P - J - K I > 240 + 2 and I (V (P) - V (P')) n G' I < 2. So, by (8), for any set Y of less than 0 vertices of G' there is a component of G- Y - P containing at least 160 + 3 vertices of Z-P. It follows that the tangle T' = {B I B is a connected subgraph of G'-P containing more than two thirds of the vertices of Z - P} has order at least 0. Furthermore, this tangle has order at most 80 + 2 as any set of 80 + 2 vertices of Z - P is a hitting set for it. To complete the proof, we need to show that T' is indistinguishable from T in G. To this end, consider any set Y of less than ord(T') < 80 + 2 vertices of G. Clearly, Y U (fT(Y) fl Z) separates Z from T.
Since Z is T-linked, it follows that Z fl fT(Z) > IZI - (80 + 2) > 160 + 3. Thus fT (Y) = fy(Y), as required. This completes the proof of Lemma 7.8 and hence of Theorem 1.2
This proof can easily be strengthened to obtain Theorem 4.2. We omit the details because even the most doughty reader must be suffering from fatigue.
Acknowledgements
I would like to thank Colin McDiarmid for his valuable suggestions on an earlier draft. His long-term support for my project to write a survey on this topic has been even more valuable. I would also like to thank Rosemary Bailey for her painstaking work as editor. Her contributions include, but are not limited to, drawing the figures and verifying and amending the reference list.
References [1] N. Alon, P. Seymour & R. Thomas, A separator theorem for graphs with an excluded minor and its applications, in Proceedings of the 22nd Annual Association for Computing Machinery Symposium on Theory of Computing, ACM Press, New York (1990), pp. 293-299.
[2] N. Alon, P. D. Seymour & R. Thomas, Planar separators, SIAM Journal on Discrete Mathematics, 7 (1994), 184-193.
[3] D. Archdeacon & P. Huneke, A Kuratowski theorem for nonorientable surfaces, Journal of Combinatorial Theory, Series B, 46 (1989), 173-231.
[4] S. Arnborg, D. G. Corneil & A. Proskurowski, Complexity of finding embeddings in a k-tree, SIAM Journal on Algebraic and Discrete Methods, 8 (1987), 277-284.
Brambles And Tree Decompositions [5]
159
S. Arnborg, B. Courcelle, A. Proskurowski & D. Seese, An algebraic theory
of graph reduction, Technical report LaBRI-90-02, Universite de Bordeaux, 1990. [6] S. Arnborg, J. Lagergren & D. Seese, Easy problems for tree-decomposable graphs, Journal of Algorithms, 12 (1991), 308-340.
[7] S. Arnborg & A. Proskurowski, Linear time algorithms for NP-hard problems restricted to partial k-trees, Discrete Applied Mathematics, 23 (1989), 11-24.
[8] H. L. Bodlaender, Dynamic programming on graphs of bounded treewidth, in Proceedings of the 15th International Colloquium on Automata, Languages and Programming (eds. T. Lepisto & A. Salomaa), Lecture Notes in Computer Science, 317, Springer Verlag, Berlin (1988), pp. 105118.
[9] H. L. Bodlaender, Polynomial algorithms for graph isomorphism and chromatic index on partial k-trees, Journal of Algorithms, 11 (1990), 631-643.
[10] H. L. Bodlaender, A linear time algorithm for finding tree decompositions of small treewidth, in Proceedings of the 25th Annual Association for Computing Machinery Symposium on Theory of Computing, ACM Press, New York (1993), 226-234.
[11] H. Bodlaender & T. Kloks, Better algorithms for pathwidth and treewidth of graphs, in Proceedings of the 18th International Colloquium on Automata, Languages and Programming (eds. J. Leach Albert, B. Monien & M. Rodriguez Artalejo), Lecture Notes in Computer Science, 510, Springer Verlag, Berlin (1991), 544-555. [12] F. R. K. Chung, Spectral Graph Theory, American Mathematical Sociey, Providence, Rhode Island (1997). [13] B. Courcelle, The monadic second order logic of graphs. I. Recognizable sets of finite graphs, Information and Computation, 85 (1990), 12-75. [14] I. Defter & V. Neumann-Lara, Unboundedness for generalized odd cycle
traversability and a Gallai conjecture, paper presented at the Fourth Caribbean Conference on Computing, Puerto Rico, 1985.
[15] P. Erdos & L. Posa, On independent circuits contained in a graph, Canadian Journal of Mathematics, 17 (1965), 347-352.
[16] M. R. Fellows & M. A. Langston, Nonconstructive advances in polynomial-time complexity, Information Processing Letters, 26 (1987), 157-162.
B. A. Reed
160
[17] M. R. Fellows & M. A. Langston, Nonconstructive tools for proving polynomial-time decidability, Journal of the Association for Computing Machinery, 35 (1988), 727-739.
[18] I. S. Filotti, G. L. Miller & J. Reif, On determining the genus of a graph in O(IVI°(9)) steps, in Proceedings of the 11th Annual Association for Computing Machinery Symposium on Theory of Computing, ACM Press, New York (1979), pp. 27-37. [19] T. Gallai, Problem 6, in Proceedings Colloquium Tihany, Academic Press, (1966), page 362.
[20] A. Gyarfas, Fruit salad, manuscript. [21] T. C. Hu, Integer Programming and Network Flows, Addison-Wesley, Don Mills, Ontario (1969).
[22] R. M. Karp, On the complexity of combinatorial problems, Networks, 5 (1975), 45-68. [23] J. B. Kruskal, Well-quasi-ordering, the tree theorem, and Vazsonyi's conjecture, Transactions of the American Mathematical Society, 95 (1960), 210-225. [24] K. Kuratowski, Sur le probleme des courbes gauches en topologie, Fundamenta Mathematicae, 15 (1930), 271-283.
[25] J. Lagergren, Upper bounds on the sizes of obstructions and intertwines, manuscript.
[26] R. J. Lipton & R. E. Tarjan, A separator theorem for planar graphs, SIAM Journal on Applied Mathematics, 36 (1979), 177-189. [27] L. Lovasz & A. Schrijver, personal communication.
[28] J. Lynch, The equivalence of theorem proving and the interconnection problem, Association for Computing Machinery's Special Interest Group on Design Automation Newsletter, 5 1976. [29] W. McCuaig, Intercyclic digraphs, in Graph Structure Theory (Proceedings of the AMS-IMS-SIAM Joint Summer Research Conference on Graph Minors, Seattle, 1991) (eds. N. Robertson & P. Seymour), Contemporary Mathematics, 147, American Mathematical Society, Providence, Rhode Island (1993), 203-247. [30] K. Menger, Zur allgemeinen Kurventheorie, Fundamenta Mathematicae, 10 (1927), 96-115.
Brambles And Tree Decompositions
161
[31] C. St.J. A. Nash-Williams, On well-quasi-ordering infinite trees, Proceedings of the Cambridge Philosophical Society, 61 (1965), 697-720.
[32] B. A. Reed, Finding approximate separators and computing tree width quickly, in Proceedings of the 24th Annual Association for Computing Machinery Symposium on Theory of Computing, ACM Press, New York (1992), 221-228.
[33] B. Reed, Mangoes and blueberries, manuscript.
[34] B. Reed, Disjoint connected paths: faster algorithms and shorter proofs, manuscript.
[35] B. Reed, N. Robertson, P. Seymour & R. Thomas, On packing directed circuits, Combinatorica, in press. [36] N. Robertson & P. D. Seymour, Graph Minors. II. Algorithmic aspects of tree-width, Journal of Algorithms, 7 (1986), 309-322.
[37] N. Robertson & P. D. Seymour, Graph Minors. IV. Tree-width and wellquasi-ordering, Journal of Combinatorial Theory, Series B, 48 (1990), 227-254.
[38] N. Robertson & P. D. Seymour, Graph Minors. V. Excluding a planar graph, Journal of Combinatorial Theory, Series B, 41 (1986), 92-114. [39] N. Robertson & P. D. Seymour, Graph Minors. VII. Disjoint paths on a surface, Journal of Combinatorial Theory, Series B, 45 (1988), 212-254. [40] N. Robertson & P. D. Seymour, Graph Minors. VIII. A Kuratowski theorem for general surfaces, Journal of Combinatorial Theory, Series B, 48 (1990), 255-288.
[41] N. Robertson & P. D. Seymour, Graph Minors. X. Obstructions to treedecomposition, Journal of Combinatorial Theory, Series B, 52 (1991), 153-190.
[42] N. Robertson & P. D. Seymour, Graph Minors. XIII. The disjoint paths problem, Journal of Combinatorial Theory, Series B, 63 (1995), 65-110.
[43] N. Robertson & P. D. Seymour, Graph Minors. XVI. Excluding a nonplanar graph, manuscript. [44] N. Robertson & P. D. Seymour, Graph Minors. XX. Wagner's Conjecture, manuscript, 1988.
B. A. Reed
162
[45] N. Robertson, P. D. Seymour & R. Thomas, A survey of linkless embeddings, in Graph Structure Theory (Proceedings of the AMS-IMS-SIAM Joint Summer Research Conference on Graph Minors, Seattle, 1991) (eds. N. Robertson & P. Seymour), Contemporary Mathematics, 147, American Mathematical Society, Providence, Rhode Island (1993), 125-136.
[46] N. Robertson, P. Seymour & R. Thomas, Quickly excluding a planar graph, Journal of Combinatorial Theory, Series B, 62 (1994), 323-348. [47] P. Seymour, A bound on the excluded minors for a surface, manuscript.
[48] P. D. Seymour & R. Thomas, Graph searching and a min-max theorem for tree-width, Journal of Combinatorial Theory, Series B, 58 (1993), 22-33.
[49] R. Thomas, A Menger-like property of tree-width: the finite case, Journal of Combinatorial Theory, Series B, 48 (1990), 67-76.
[50] C. Thomassen, On the presence of disjoint subgraphs of a specified type, Journal of Graph Theory, 12 (1988), 101-111. [51] W. Tutte, Algebraic Theory of Graphs, Ph. D. Thesis, Cambridge, 1948. [52] V. G. Vizing, On the estimate of the chromatic class of a p-graph, Metody Diskretnogo Analiza, 3 (1964), 25-30.
[53] D. Younger, Graphs with interlinked directed circuits, in Proceedings of the Midwest Symposiom on Circuit Theory (1973), pp. XVI 2.1-2.7.
Equipe Combinatoire CNRS Case 189 Universite de Paris VI 4 place Jussieu Paris 75005 France
Minor-monotone Graph Invariants Alexander Schrijver Summary A graph parameter q5(G) is called minor-monotone if ¢(H) < ¢(G) for any minor H of G. We survey recent work on minor-monotone graph parameters motivated by the parameter p(G) introduced by Colin de Verdiere. 1
Introduction A function q5(G) defined for any undirected graph G is called minor-monotone
if for any graph G and any minor H of G one has q5(H) < O(G)
In this paper, all graphs are undirected, loopless and without multiple edges. A minor of a graph arises by a series of deletions and contractions of edges and deletions of isolated vertices, suppressing any multiple edges and loops that may arise. The interest in minor-monotone graph parameters is activated because the Robertson-Seymour theory of graph minors can be applied to them. Recently a number of minor-monotone parameters have been studied, motivated in particular by the graph parameter p(G) introduced by Colin de Verdiere [5] (cf. [6]). The parameter µ(G) can be described in terms of properties of matrices related to G. It was motivated by the study of the maximum multiplicity of the second eigenvalue of certain Schrodinger operators. When such an operator is defined on a Riemann surface, one can approximate the surface by a densely enough embedded graph G, in such a way that p(G) is the maximum multiplicity of the second eigenvalue of the operator. The interest raised by Colin de Verdiere's parameter can be explained not only by its background in differential geometry, but also by the facts that it is minor-monotone (so that the Robertson-Seymour graph minors theory applies to it), and that it characterizes planarity of graphs. Indeed, one has that µ(G) < 3 if and only if G is planar. Moreover, as follows from the results in [19] and [15], p(G) < 4 if and only if G is linklessly embeddable in R3. (A graph G is linklessly embeddable if it can be embedded in R3 in such a way that the images of any two disjoint circuits in G are unlinked.) So with the help of p, topological properties of a graph can be characterized in terms of spectral properties of matrices associated to the graph. In this paper we give a survey of the graph parameter µ(G), and some related parameters, in particular the parameter A(G) introduced in [12]. We first give an overview of µ(G) and A(G), after which we give proofs of a number of results. Finally, we consider the parameters )'(G) (defined by oriented matroids) and ic(G). 163
Alexander Schrijver
164
For more information we refer to the thesis by van der Holst [11], where in addition a few other minor-monotone parameters are studied. 2
Overview of µ(G)
Let G = (V, E) be an undirected graph, which we assume without loss of generality to have vertex set V = {1, . . . , n}. Then µ(G) is the largest corank of any symmetric real-valued n x n matrix M = (mij) such that: (i) M has exactly one negative eigenvalue, of multiplicity 1,
(2)
(ii) for all i, j with i : j: mij < 0 if i and j are adjacent, and mi,j = 0 if i and j are nonadjacent,
(iii) there is no nonzero symmetric n x n matrix X = (xij) such that MX = 0 and such that xij = 0 whenever i = j or mi,j # 0. There is no condition on the diagonal entries mi,i. The corank corank(M) of a matrix M is the dimension of its kernel (= null space). Note that for each graph G = (V, E) a matrix M satisfying (2) exists. If G is connected, let A be the adjacency matrix of G. Then we can choose A in such a way that .\I - A has exactly one negative eigenvalue and is nonsingular. If G is disconnected, we can choose such a A for each component separately and obtain again a nonsingular matrix with exactly one negative eigenvalue. Condition (iii) is called the Strong Arnol'd Hypothesis (or Strong Arnol'd Property). There are a number of equivalent formulations of the Strong Arnol'd Hypothesis, amounting to the fact that M is in a certain general position. Let
M = (mij) be a symmetric n x n matrix. Let RM be the set of all symmetric n x n matrices A with rank(A) = rank(M). Let SM be the set of all symmetric n x n matrices A = (aij) such that aij = 0 whenever i # j and mi,j = 0. Then M fulfils the Strong Arnol'd Hypothesis (2)(iii) if and only if RM intersects SM at M `transversally';
(3)
that is, if the tangent space of RM at M and the tangent space of SM at M together span the space of all symmetric n x n matrices. In other words, if the intersection of the normal spaces at M of RM and of SM only consists of the all-zero matrix. It is elementary linear algebra to show that the tangent space of RM at M
consists of all symmetric n x n matrices N such that xT Nx = 0 for each x E ker(M). Thus the normal space of RM at M is equal to the space generated by all matrices xxT with x E ker(M). (We assume that our underlying space is the space of real-valued symmetric n x n matrices.) This space is equal to the space of all symmetric n x n matrices X satisfying MX = 0. Trivially, the
Minor-monotone Graph Invariants
165
normal space of SM at M consists of all symmetric n x n matrices X = (xi,,j) such that xij = 0 whenever i = j or mij 0. Therefore, (3) is equivalent to (2) (iii).
An important property of µ(G) proved by Colin de Verdiere [5] is that it is monotone under taking minors: the graph parameter µ(G) is minor-monotone.
(4)
Proving this is nontrivial, and the Strong Arnol'd Hypothesis is needed. We give the elementary proof as given in van der Holst [11] in Section 4. The minor-monotonicity of µ(G) is especially interesting in the light of the Robertson-Seymour theory of graph minors [16], which has as principal result that if C is a collection of graphs so that no graph in C is a minor of another graph in C, then C is finite. This can be equivalently formulated as follows. For any graph property P closed under taking minors, call a graph G a forbidden minor for P if G does not have property P, but each proper minor of G does have property P. Note that a minor-closed property P is completely characterized by the collection of its forbidden minors. Now Robertson and Seymour's theorem states that each graph property that is closed under taking minors, has only finitely many forbidden minors. (See Reed's paper elsewhere in this volume.) Since
p(Kn)=n-1
(5)
for each n (cf. Section 5), Hadwiger's conjecture implies that y(G) < p(G) + 1 (where y(G) denotes the colouring number of G); this last inequality is conjectured by Colin de Verdiere [5]. Since Hadwiger's conjecture holds for graphs not containing any K6-minor (Robertson, Seymour, and Thomas [18]), we know
that y(G) < µ(G) + 1 holds if µ(G) < 4. In studying p(G), we can restrict ourselves to considering connected graphs,
since if G has at least one edge, then p(G) is equal to the maximum of µ(K) taken over all components K of G. The following characterizations show that with the help of µ(G), topological properties of a graph can be characterized algebraically: (i) µ (G)
n - 3 then G is outerplanar; if u(G) > n - 4 then G is planar.
(8)
The proof by Colin de Verdiere [5] of the planarity characterization (6) (iii) uses a result of Cheng [4] on the maximum multiplicity of the second eigenvalue of Schrodinger operators defined on the sphere. A short direct proof was given
by van der Holst [10], based on the following lemma. For any vector x, let supp(x) denote the support of x (i.e., the set {i I xi 0}). Moreover, denote supp+(x) := {i I xi > 0} and supp-(x) := {i I xi < 0}. We say that a vector x E ker(M) has minimal support if x is nonzero and for each nonzero vector
y E ker(M) with supp(y) C supp(x) one has supp(y) = supp(x). For any subset U of V, let GIU denote the subgraph of G induced by U. Then Van der Holst's lemma states:
Let M satisfy (2) and let x E ker(M) have minimal support. Then GIsupp+(x) and GIsupp-(x) are connected.
(9)
We give the proof in Section 10. 3
Overview of )t(G)
Van der Hoist's lemma motivated van der Holst, Laurent, and Schrijver [12] to introduce a related graph parameter )(G), defined as follows. Let G = (V, E) be a graph. Call a subspace X of W" representative for G if
for each nonzero vector x E X, supp+(x) is nonempty and GI supp+(x) is connected.
(10)
Minor-monotone Graph Invariants
167
Then A(G) is defined as the maximum dimension of any representative subspace X of llly . Clearly, (10) implies that also supp-(x) is nonempty and induces a connected subgraph of G for each nonzero x E X. The results characterizing p and A for small values, suggest that A is close to p. In fact, recently Rudi Pendavingh showed that µ(G) < A(G) + 2 for each graph G. Conversely, it might be that A(G) < µ(G) holds. There is a direct equivalent characterization of A(G). Let G = (V, E) be a graph and let d E N. Call a function 0: V -* Rd representative for G if
for each halfspace H of lRd, the set 0-1(H) is nonempty and induces a connected subgraph of G.
(11)
(Here 0-1(H) := {v E V I 0(v) E H}.) A subset H of Rd is called a halfspace if H = {x E Rd I cTx > 0} for some nonzero c E W. Note that if 0: V -+ Rd is representative, then the vectors 0(v) (v E V) span Rd (since otherwise there would exist a halfspace H with ¢-1(H) = 0). Now A(G) is equal to the largest d for which there is a representative function 0: V -> Rd. This is easy to see. Suppose X is a d-dimensional subspace of ]R' representative f o r G. Let vectors x1, ... , xd form a basis of X. Define ¢(v) := (X1 (V), ... , xd(v)) for each v E V. Then 0 is a representative function for G. Conversely, let 0: V -* Rd be representative. Define for any
c E Rd the function x, E IWY by: x,(v) := cT¢(v) for v E V. Then X {x, I c E Rd } is a representative space for G. It is easy to show that the function A(G) is minor-monotone (much easier
than for µ(G)):
Theorem 3.1 If H is a minor of G then A(H) < A(G). Proof Let H = (V', E'). If H arises from G by deleting an isolated vertex vo, the inequality A(H) < A(G) is easy: if ¢': V' -* Rd is representative for H with d = A(H), then defining ¢(vo) := 0 and q5(v) := q'(v) for all other vertices v of G, gives a representative function for G. So we may assume that H = (V', E') arises from G = (V, E) by deleting or contracting one edge e = uw. Let q': V' -* Rd be representative for H with
d = A(H). If H arises from G by deleting e, then V = V', and 0' is also representative for G. Hence A(G) > d = A(H). If H arises from G by contracting e, let vo be the vertex of H which arose by contracting e. Define 0(u) := q5(w) :_ 4'(vo), and define 0(v) := 4'(v) for all other vertices v of G. Then 0 is representative of G. One easily shows that
A(K,,) = n - 1
(12)
(cf. Section 5). Hence, Hadwiger's conjecture implies that 'y(G) < A(G) + 1 (where 'y(G) denotes the colouring number of G). So by the truth of Hadwiger's
Alexander Schrijver
168
conjecture for K6-free graphs (Robertson, Seymour, and Thomas [18]), the inequality y(G) < A (G) + 1 holds if A (G) < 4. As for the colouring number, also the function A (G) cannot be increased by `clique sums'. Graph G = (V, E) is a clique sum of graphs G1 = (V1, E1) and G2 = (V2, E2) if V = V1 U V2 and E = E1 U E2, where V1 fl V2 is a clique
both in G1 and in G2. Then y(G) = max{y(G1), y(G2)} if G is a clique sum of Gl and G2. A similar relation holds for the size of the largest clique minor in G. Now in Section 6 we shall show: If G has at least one edge and is a clique sum of G1 and G2,
(13)
then A(G) = max{A(Gi), A(G2)}.
This directly gives with (12):
(i) A(G) < 1 if and only if G is a forest;
(14)
(ii) ) (G) < 2 if and only if G is a series-parallel graph. Indeed, forests can be characterized as the graphs not having a K3-minor and also as the graphs obtainable from K2 by taking clique sums and subgraphs. Similarly, series-parallel graphs can be characterized as the graphs not having a K4-minor and also as the graphs obtainable from K3 by taking clique sums and subgraphs. In Section 13 we show that A (G) < 3 if and only if G can be obtained from planar graphs by taking clique sums and subgraphs.
(15)
The kernel of the proof here is to show that A (G) < 3 for any planar graph G. Having this, a fundamental decomposition theorem of Wagner [20] then implies the full characterization. Indeed, let V8 be the graph with vertices v1, ... , v8,
Figure 1: The graph V8
where vi and vj are adjacent if and only if I i - j I E {1,4,7}. Then Wagner
Minor-monotone Graph Invariants
169
showed:
G can be obtained from planar graphs by taking clique sums and subgraphs G does not have a K5- or V8-minor.
(16)
Since )(K5) = 4 and since A(V8) = 4 (as we show in Section 13), we obtain (15). In Section 16 we give a few observations concerning the class of graphs G with A(G) < 4. In particular, we show the result of [15] that A(G) < 4 for any linklessly embeddable graph G. This implies with (13):
if G is obtainable from linklessly embeddable graphs by taking clique sums and subgraphs, then A(G) < 4.
(17)
As mentioned, an open question is if there is any direct relation between
X(G) and p(G). It might be the case that )(G) < µ(G) for each graph G. That is, for any subspace X of R" representative for G there is a matrix M satisfying (2) with dim(X) < corank(M). This is true if µ(G) < 4. In fact, a tempting, more general speculation is that for any natural number t: (???) a graph G satisfies A(G) < t if and only if G is obtainable from graphs H satisfying µ(H) < t by taking clique sums and subgraphs (???)
(18)
This has been proved for t < 3, and the `if' part for t < 4. 4
Some basic facts on µ(G)
We first prove a number of elementary facts on the parameter µ(G). We use the following notation. If M is a matrix and I is a set of rows of M and J is a set of columns of M, then Mj,< j is the submatrix induced by the rows
in I and the columns in J. If I = J we write MI for Mlr First we have the following important property due to Colin de Verdiere [6], which we prove with the method described by van der Holst [11]:
Theorem 4.1 For any edge e of any graph G one has p(G - e) < µ(G). Proof For any smooth manifold M, any smooth submanifold A of Rd, any smooth function f : M -> Rd, and any x E M with f (x) E A, we say that f intersects A transversally at x E M, in notation: f #xA, if Tj(x)A+Dff(TTM) = Rd.
(19)
Here TAN denotes the tangent space of N at y, and D ff the differential of f at x.
Alexander Schrijver
170
A basic property of transversality is:
If f #xA, then there is a neighbourhood U of x in M fl f -1(A) (20) such that f #yA for each y E U, and such that U has the same codimension in M as A has in 1[ld. Let Sn denote the collection of real-valued symmetric n x n matrices, and Sn,k the collection of matrices in Sn of corank k. For any graph G = (V, E) let OG be the collection of real-valued symmetric V x V matrices M satisfying (2) (ii). First assume that graph H arises from graph G by deleting an edge e = uw.
We may assume that G has vertex set V = {1,.. . , n}, and that u = 1 and w = 2. Let W :_ {3, ... , n}. Let f : R x OH -4 Sn be defined by
f (h, K) :=
k1,1
h
K{i}xw
h
k2,2
K{2}xw
(21)
Kwx{1} Kwx{2} Kwxw where K = (ki,j) E OH. Let f0(K) := f (0, K). Let M= (mid) satisfy (2), k = µ(H). By (2)(iii), fO#M'Sn,k,
(22)
f #(o,M')Sn,k
(23)
which implies
Then by (20) there is a neighbourhood U of (0, M') in R x OH such that for all x E U f #xSn,k
(24)
Also by (20), u fl ({0} x fo 1(Sn,k)) is a submanifold of u fl ({0} x OH) of codimension 2k(k + 1) (since the codimension of Sn,k in Sn is Zk(k + 1)). Moreover, f-1(Sn,k) fl u is a submanifold of U of codimension Zk(k + 1). Hence there exists a (h, L) E U with h < 0 such that M := f (h, L) E Sn,k. By taking (h, L) close to (0, M') we may assume that M has exactly one negative eigenvalue. Since f #(h,L)Sn,k, M fulfils the Strong Arnol'd Hypothesis ((3)). Hence M satisfies (2), and therefore p(G) > µ(H). This theorem implies:
Theorem 4.2 For any subgraph H of any graph G one has
p(H) < p(G).
(25)
Proof By Theorem 4.1 we can assume that H arises from G = (V, E) by deleting an isolated vertex v. Let M' be a matrix satisfying (2) with respect
to H, with corank(M') = p(H), and let M be the V x V matrix arising from M' by adding 0's, except in position (v, v), where M,,,v = 1. Then trivially corank(M) = corank(M') and M satisfies (2) with respect to G. This shows (25).
Minor-monotone Graph Invariants
171
This implies:
Theorem 4.3 If G has at least one edge, then p(G) = mKax p (K),
(26)
where K extends over the components of G. Proof By Theorem 4.2 we know that > holds in (26). To see equality, let M be a matrix satisfying (2). Since G has at least one edge, we know µ(G) > 0 (since
trivially µ(K2) = 1), and hence corank(M) > 0. Then there is exactly one component L of G with corank(ML) > 0. For suppose that there are two such components, K and L. Choose nonzero vectors x E ker(MK) and y E ker(ML). Extend x and y by zeros on the positions not in K and L, respectively. Then the matrix X := xyT + yxT is nonzero and symmetric, has zeros in positions corresponding to edges of G, and satisfies MX = 0. This contradicts the Strong Arnol'd Hypothesis. So corank(M) = corank(ML). Suppose now that ML has no negative eigenvalue. Then 0 is the smallest eigenvalue of ML, and hence, by the connectivity of L and the Perron-Frobenius theorem, corank(ML) = 1. So µ(G) = 1. Let L' be a component of G with at least one edge. Then µ(L') > 1, proving (26). One easily shows that ML satisfies the Strong Arnol'd Hypothesis, implying µ(G) = µ(L), thus proving (26). Next we have:
Theorem 4.4 Let G = (V, E) be a graph and let v E V such that G - v has at least one edge. Then p(G) < µ(G - v) + 1.
(27)
Proof Let M be a matrix satisfying (2) with corank(M) = µ(G). Let M':= My\{v} Clearly, corank(M') > corank(M) - 1, since rank(M) < rank(M). So it suffices to show that M' satisfies (2) with respect to G'. Trivially, M' satisfies (2) (ii). To see that M' satisfies (2) (i), it suffices to show that M' has at least one negative eigenvalue. If M' has no negative eigenvalue, then M' is positive semidefinite, and 0 is an eigenvalue of multiplicity at least p(G) + 1. Hence (by the Perron-Frobenius theorem) for each component K of G - v, if the matrix MK has eigenvalue 0, then it has multiplicity 1. As the theorem trivially holds if µ(G) < 2 (since p(G - v) > 1 as G - v has
at least one edge), we can assume that p(G) > 3. Hence G - v has at least µ(G) + 1 > 4 components K with MK singular. Let K1,. .. , K4 be four such components. For i = 1, ... , 4, let xi be a nonzero vector with MK:xi = 0. By the Perron-Frobenius theorem we know that we can assume xi > 0 for each i. Extend xi to a vector in iW' by adding components 0.
Alexander Schrijver
172
Let z be an eigenvector of M belonging to the smallest eigenvalue of M. By scaling the xi we can assume that zT xi = 1 for each i. Now define X := (xl - x2)(x3 - X4 )T + (x3 - x4)(xl - X2 )T.
(28)
Then MX = 0, since M(xl - x2) = 0 (as (x1 - x2)TM(xl - x2) = 0 and as xl - x2 is orthogonal to z), and similarly M(x3 - x4) = 0. This contradicts the fact that M satisfies (2)(iii). So M' satisfies (2)(i). To see that M' satisfies the Strong Arnol'd Hypothesis (2) (iii), let X' be a (V \ v) x (V \ v) matrix with 0's in positions (i, j) where i = j or i and j are adjacent, and satisfying M'X' = 0. We must show that X' = 0. Let X be the V x V matrix obtained from X' by adding 0's. Since M' has exactly one negative eigenvalue, we know by interlacing that
corank(M') < corank(M). If MX = 0 we know by (2)(iii) that X = 0 and hence X' = 0. So we can assume that MX
0. As corank(M') < corank(M), it follows that there is a vector x E ker(M) with x, # 0. Hence the first column of M is a linear combination of the other columns of M. Therefore MX = 0, a contradiction. On the other hand we have, where S(G) arises from G by adding one new vertex v adjacent to all other vertices of G:
Theorem 4.5 For any graph G with at least one edge, one has p(S(G)) = µ(G) + 1.
(29)
Proof By Theorem 4.4 it suffices to show that p(S(G)) > p(G) + 1, and by Theorem 4.3 we can assume that G is connected. Let M be a matrix satisfying (2) with corank(M) = µ(G). Let z be an eigenvector of M belonging
to the smallest eigenvalue Al of M. We can assume that z < 0 and that lizil = 1. Let M' be the matrix
/
'
M' := I\ z l M zT 1
(30)
Since (0, x)T E ker(M') for each x E ker(M) and since (-Al, z)T E ker(M'), we know that corank(M') > corank(M) + 1. By interlacing it follows that M' has exactly one negative eigenvalue. One similarly easily checks that M' satisfies the Strong Arnol'd Hypothesis (2)(iii).
Above we gave a proof that a(G) is monotone under taking subgraphs. More strongly, as Colin de Verdiere [5] proved, µ(G) is minor-monotone. Again we give the elementary proof due to van der Holst [11] of this fact.
Theorem 4.6 µ(G) is minor-monotone.
Minor-monotone Graph Invariants
173
Proof By Theorem 4.2 it suffices to show that µ(H) < µ(G) if H arises from G by contracting edge e = uw. Let the new vertex of H be v. Let n IVI and n' 1V'I. Son = n' + 1. We may assume that u = 1 and w = 2. Let {3, ... , n}. W
Let 2 be the set of all matrices K = (ki,j) E OG with k1,1 = 0 = k1,2. Define a function
f:1RxZ-4Sn,
(31)
by
K{1}xw + K{2}xw
k2,2
f (h K) =
(32)
Kwx{1} + Kwx{2} Kwxw - hKwx{1}K{1}xw
and let fo(K) = f (0, K).
Let M' = (m'ij) satisfy (2) with respect to H, with corank k = µ(H). Trivially there is a P E Z such that f (0, P) = M'. Since the tangent space of OH at M' is a subspace of the space of all vectors Df(o,P)(A) with A E T(o,p)(R x Z) we know that f #(o,P)Sn',k
(33)
Again by (20), there is a neighbourhood U of (0, P) such that for all x E U (34)
f #xSn,,k.
Also by (20), ({0} x fo 1(Sn',k)) fl U is a submanifold of U fl ({0} x OH) of codimension 2k(k + 1) (since the codimension of Sn',k in S,,, is 2k(k + 1)). Moreover, f -1(Sn,,k) n U is a submanifold of U of codimension k(k + 1). Hence there is an (h, L) E U with h > 0 such that f (h, L) E Sn',k and f #(h,L)Sn',k
(35)
Taking (h, L) close to (0, P) we may assume that f (h, L) has exactly one negative eigenvalue. Define
L{l}xw
M:=
)
L{2}xw Lwx{1}
Lwx{2}
(36)
Lwxw )
Clearly M E OG. We show that M satisfies (2) and has corank k. Let
P:=
1
1
-hL{l}xw
0
1
0
0
0
I
(37)
Alexander Schrijver
174
Then 1
PT MP =
h
0
0
f (h, L)
(38)
Therefore, by Sylvester's law of inertia and since h > 0, f (h, L) has the same number of negative eigenvalues and the same corank as M. It remains to show that M fulfils the Strong Arnol'd Hypothesis ((2)(iii)). Choose F E S. We must show that there exists an N E TMOG such that xTFx = xTNx for all x E ker(M). Define 1
-hL{1}xw
1
0
0
I
(39)
and F' := QTFQ. Since f #(h,L)Sn',k, Df(h,L) (T(h,L) (R x Z)) + Tf(h,L) (Sf',k) = Sn'
(40)
The tangent space of Sn" k at f (h, L) is the set of all real-valued symmetric matrices C for which x'TCx' = 0 for all x' E ker(f (h, L)). Hence there is an (a, B) E T(h,L) (IR x Z) such that
xTDf(h,L)(a,B)x -xTF'x
(41)
for all x' E ker(f (h, L)). Now let
N :_
a
a
h2
h2
a - h2
B(2}.121 +
BWx{1}
BWx{2}
B{l}xw
a h2
B{2} x W
(42)
BWxw
So N E TMON. A calculation shows
Df(h,L)(a,B) =QTNQ.
(43)
For each vector x E ker(M), the vector X/
=
x2
(44)
xWy
belongs to ker(f (h, L)) and satisfies Qx' = x. Hence xTFx = x'T QT FQx' = x'T D f(h,L) (a, B)x' = xTNx.
(45)
Minor-monotone Graph Invariants 5
175
µ(G) and A(G) for complete graphs It is easy to see that for each graph G with n vertices one has
p(G) 0} containing Z. (Thus cone(O) = {0}, while cone(Z) = Rd if there are no halfspaces containing Z.) For any graph G = (V, E) and U C V, let G - U denote the graph obtained from G by deleting the vertices of U. (So G - U = GI(V \ U).)
Theorem 6.1 Let 0: V -4 Rd be representative for a graph G = (V, E) and let U C V. Assume that cone(O(U)) is not a hyperplane in Rd. Then there is at most one component K ofG-U for which the inclusion O(K) C cone(O(U)) does not hold.
Proof We may assume that cone(O(U)) # Rd. Since cone(q5(U)) is not a hyperplane in Rd, the set
C :_ {c E R' I c 54 0, cT cb(v) < 0 for each v E U},
(47)
is nonempty and topologically connected (because the polar cone C U {0} of
cone(q(U)) is not a line). For c E Rd, let H.:= {x E Rd I cTx > 0}. Let K1i . . . , Kt be the components of G - U. Let Ci be the set of vectors c E C for which HH intersects q (Ki). So if i # j then Ci fl C; = 0, since if c E C then 0-1(He) is connected and is disjoint from U. As C1 U ... U Ct = C and since each Ci is an open subset of C, it follows that Ci = 0 for all but one i. Hence O(Ki) C cone(q(U)) for all but one i. This implies ([12]):
Theorem 6.2 If G has at least one edge and is a clique sum of Gl and G2, then A(G) = max{a(Gi), ) (G2)}.
(48)
Proof We have A(G) > max{A(Gi), A(G2)}, since G1 and G2 are subgraphs of G, So it suffices to show that A(G) = .(Gi) for some i = 1, 2. Assume that A(G) > max{A(Gi), A(G2)}. Let d:= A(G), G = (V, E), and Gi = (V, Ei) for
i=1,2.
Minor-monotone Graph Invariants
177
Let ¢: V -* Rd be representative for G. As d > )(Gi), Ol V is not represent-
ative for Gi, for i = 1 and i = 2. Let K := Vi n V2 and t := JKI. We may assume that we have chosen the counterexample so that IKI is as small as possible. Then GI (V1 \ K) has a component L such that each vertex in K is adjacent
to at least one vertex in L. Otherwise G would be a repeated clique sum of subgraphs of G1 and G2 with common clique being smaller than K. In that case A(G) = max{.\(G1),.\(G2)} would follow by the minimality of K. So G1 has a Kt+i-minor. So \(G1) > t, and hence A(G) > t = I KI. Therefore, cone(O(K)) is not a hyperplane in R. (Here we use that it is not the case that K = 0 and d = 1.) So by Theorem 6.1, we may assume that O(V1) C cone(O(K)). As d > .\(G2), there exists a halfspace H of Rd such that GI (q-1(H) n V2)
is empty or disconnected. If it is empty, then ¢(v) E H for some v E V1 \ K, contradicting the facts that 0(v) E cone(qb(K)) and that ¢(K) n H = 0. So it is disconnected. But then also ¢-1(H) would induce a disconnected subgraph of G, as K is a clique. This is a contradiction. Hence we have that for each t > 1:
the class of graphs G with A(G) < t is closed under taking
(49)
clique sums.
A statement like this for p does not hold. A critical example is the graph Kt+3 \ A (the graph obtained from the complete graph Kt+3 by deleting the
edges of a triangle). One has p(Kt+3 \ A) = t + 1 (since the star K4 \ A has µ(K4 \ A) = 2 (see Theorem 8.2 below), and since adding a new vertex adjacent to all existing vertices increases p by 1). However, Kt+3 \ A is a clique sum of Kt+1 and Kt+2 \ e (the graph obtained from Kt+2 by deleting an edge), with common clique of size t. Both Kt+1 and Kt+2 \ e have p = t. So, generally one does not have that, for fixed t, the property p(G) < t is maintained under clique sums. Similarly, Kt+3 \ A is a clique sum of two copies of Kt+2 \ e, with common clique of size t + 1. These examples where µ increases by taking a clique sum are in a sense the only cases, as shown in [13]:
Theorem 6.3 If G has at least one edge and is a clique sum of G1 and G2, with common clique S, then p(G) > t := max{µ(G1), p(G2)} if and only if. either (i) ISI = t and G - S has three components the contraction of which makes with S a Kt+3 \ 0,
(50)
or (ii) I S I = t+l and G-S has two components the contraction of which makes with S a Kt+3 \ A-
Moreover, if p(G) > t then p(G) = t + 1, µ(G1) = µ(G2) = t, and we can contract two or three components of G - S so that the contracted vertices together with S form a Kt+3 \ 0-
Alexander Schrijver
178
7
Behaviour of µ(G) and A(G) under YO and AY
The results on clique sums can be applied to study the behaviour of µ(G) and A(G) under applying the YD- and AY-operations. The YO-operation works as follows, on a graph G: choose a vertex v of degree 3, make its three neighbours pairwise adjacent, and delete v and the three edges incident with v. The AY-operation is the reverse operation, starting with a triangle and adding a new vertex. Note that the if H arises by a AY from G, then H is a subgraph of a clique
sum of G and K4. Then Theorem 6.3 implies that p(H) < µ(G) if µ(G) > 4, and Theorem 6.2 that \(H) < A(G) if A(G) > 3. In fact, Bacher and Colin de Verdiere [1] proved:
Let H arise by a AY operation from G. Then p(H) < µ(G). If moreover p(G) > 4, then p(H) = µ(G).
8
(51)
p (G) and A (G) for complete bipartite graphs
Since complete bipartite graphs are often candidates for forbidden minors, in this section we give formulas for µ(Km,n) and X(Km,n). This also exhibits a difference between µ(G) and A (G). First we consider A(G):
Theorem 8.1 For n > m > 1, A(Km,n) = m. Proof On the one hand, Km+l is a minor of Km,n, and on the other hand, Km,n is a subgraph of a clique sum of Km+1's. So by Theorem 6.2 )t(Km,n) _ .\(K,n+i) = in.
Characterizing µ(G) for complete bipartite graphs is a little more complicated:
Theorem 8.2 For n > m > 1 we have µ(Km,n)=SI
m
if n < 2,
m+1 ifn>3.
(52)
Proof Note that u(Km,n) < m + 1 by Theorem 6.3, since Km,n is a subgraph of a clique sum of Km+1's. It is not hard to see that µ(K1,1) = µ(K1,2) = 1 and p(K2,2) = 2. Hence a(Km,n) = m if n < 2. So let n > 3. If m < 3 we can assume that n = 3. Let Km,3 have vertices 1,...,m+3, with colour classes {1, ... , m} and {m+ 1, m+2, m+3}.
Let M be the (m + 3) x (m + 3) matrix with mij = -1 if i < m < j or j < m < i, and mij = 0 otherwise. Then M has rank 2 and hence corank m+1. Moreover M satisfies (2). Indeed, (2)(ii) is trivial. Moreover, (2)(i) follows directly from the fact that neither M nor -M is positive semi-definite. Finally,
Minor-monotone Graph Invariants
179
M satisfies the Strong Arnol'd Hypothesis ((2)(iii)). Otherwise there is a nonzero symmetric matrix X with MX = 0 and xij = 0 if i = j or i < m < j, which can be seen to be impossible using the fact that m < n = 3. If m > 4, we can assume that n = m. Choose two adjacent vertices u and v of Km,m. Delete the edge uv, and delete m - 4 other edges incident with u and m - 4 other edges incident with v. So in the new graph, u and v have degree 3. Applying YO to u and to v we obtain a Km-,,m_1 with a triangle added to each of the colour classes. The 2(m - 4) vertices not covered by these triangles span a matching of size m - 4. Contracting each edge of this matching, we obtain a Km+2. Since A(Km+2) = m+1, we obtain ii(Km,m) = m+1 (using (51)).
9
Characterizing A (G) < 1 and µ(G) < 1 Note that one trivially has:
µ(G) = 0
.\(G) = 0
G has exactly one vertex.
(53)
We next describe the collections of graphs G satisfying µ(G) < 1 and A(G) < 1. For p(G) it is ([5]):
Theorem 9.1 µ(G) < 1 if and only if G is a vertex-disjoint union of paths; that is, if G does not have a K3 or K1,3-minor. Proof Since t.c(K3) = 2 by Theorem 5.1 and µ(K1,3) = 2 by Theorem 8.2, the minor-monotonicity of p gives the `only if' part. To see the `if' part, we can assume, by the minor-monotonicity of µ(G), that G is a path. Then trivially any matrix M satisfying (2) has rank at least n - 1, and hence corank at most 1. So ii(G) < 1. The class of graphs G with A(G) < 1 is a little larger ([12]):
Theorem 9.2 )t(G) < 1 if and only if G is a forest; that is, if and only if G does not have a K3-minor.
Proof If \(G) < 1 then G has no K3-minor, as ) (K3) = 2. Conversely, if G is a forest, then G arises by taking clique sums and subgraphs from the graph K2. As )(K2) = 1, Theorem 6.2 gives the corollary.
10
Van der Holct's lemma
In characterizing µ(G) < 2 and µ(G) < 3 the lemma due to van der Holst [10] turns out to be very helpful. If x E R" and I C 11, . . . , n}, then xl denotes the subvector of x induced by the indices in I. Recall that a vector x E ker(M) has minimal support if x is nonzero and for each nonzero vector y E ker(M) with supp(y) C supp(x) one has supp(y) _ supp(x).
Alexander Schrijver
180
Theorem 10.1 (Van der Hoist's lemma) Let G be a connected graph and let M satisfy (2). Let X E ker(M) have minimal support. Then GI supp+(x) and GI supp-(x) are both connected. Proof Suppose that (say) GI supp+(x) is disconnected. Let I and J be two of the components of GI supp+(x). Let K := supp-(x). Since mij = 0 if i E I, j E J, we have: Mrxrxr + MIxKXK = 0, (54) MJxJxJ + MJxKXK = 0.
Let z be an eigenvector of M with negative eigenvalue. By the PerronFrobenius theorem we may assume z > 0. (Strictly speaking, we apply the Perron-Frobenius theorem to the (nonnegative and indecomposable) matrix Al - M choosing A large enough.) Let zI xI z3xJ
(55)
Define y E R" by: yi := xi if i E I, yi := -Axi if i E J, and xi := 0 if i 0 I U J. By (55), zT y = zI xI - Az,Tl x J = 0. Moreover, one has (since mi, = 0
if i E I and j E J): (56)
yT My = yi Mlxryl + YTi MJxJyJ
= xI Mrxrxr +
j MJxJxJ
A2x
-XIMIxKXK - AZxJ
KXK
< 0,
(using (54)) since MIXK and MJxK are nonpositive, and since xI > 0, xj > 0 and XK < 0.
Now zTy = 0 and yT My < 0 imply that My = 0 (as M is symmetric and has exactly one negative eigenvalue, with eigenvector z). Therefore, y E ker(M). This contradicts the fact that x has minimal support.
We note that if M satisfies (2), then each vertex v ¢ supp(x) adjacent to some vertex in supp+(x) is also adjacent to some vertex in supp-(x), and conversely; that is, for each x E ker(M):
(57)
N(supp+(x)) \ supp(x) = N(supp-(x)) \ supp(x).
Here N(U) is the set of vertices in V \ U that are adjacent to at least one vertex in U.
Minor-monotone Graph Invariants
181
Characterizing µ(G) < 2 and .(G) < 2
11
We can now derive the following result of Colin de Verdiere [5]:
Theorem 11.1 µ(G) < 2 if and only if G is outerplanar; that is, if and only if G does not have a K4- or K2,3-minor.
Proof Since µ(K4) = 3 by Theorem 5.1 and µ(K2,3) = 3 by Theorem 8.2, the minor-monotonicity of µ gives the `only if' part (using the forbidden minor characterization of outerplanarity).
To see the `if' part, we may assume that G is maximally outerplanar. Suppose that µ(G) > 2, and let M be a matrix satisfying (2) of corank more than 2. Let uv be a boundary edge of G. Then there exists a nonzero vector x E ker(M) with xu = x = 0. We can assume that x has minimal support. By Van der Holst's lemma (Theorem 10.1), GI supp+(x) and GI supp-(x) are nonempty and connected. As G is maximally outerplanar, G is 2-connected. Hence there exist two vertex-disjoint paths Pl and P2 from supp(x) to {u, v}.
Let Pi and P2 be the parts outside supp(x). Then the first vertices of Pi and PP both belong to N(supp(x)), and hence (by(57)) to both N(supp+(x)) and N(supp-(x)). Contracting each of supp+(x), supp-(x), Pl, and P2 to one point, gives an embedded outerplanar graph with uv on the boundary and u and v connected by two paths of length two. This is not possible. The corresponding characterization for A (G) is easier, and was given in [12]:
Theorem 11.2 A(G) < 2 if and only if G is a series-parallel graph; that is, if and only if G does not have a K4-minor. Proof If A(G) < 2 then G has no K4-minor, as \(K4) = 3. Conversely, if G is a series-parallel graph, then G arises by taking clique sums and subgraphs from the graph K3. As A(K3) = 2, Theorem 6.2 gives the corollary.
12
Characterizing µ(G) < 3 We apply Van der Holst's lemma (Theorem 10.1) similarly to the case
< 3, a main result of Colin de Verdiere [5]:
Theorem 12.1 µ(G) < 3 if and only if G is planar; that is, if and only if G does not have a K5- or K3,3-minor.
Proof Since µ(K5) = 4 by Theorem 5.1 and µ(K3,3) = 4 by Theorem 8.2, the minor-monotonicity of µ gives the `only if' part (using Kuratowski's forbidden minor characterization of planarity).
Alexander Schrijver
182
To see the `if' part, we may assume that G is maximally planar (triangulated). Suppose that p(G) > 3 and let M be a matrix satisfying (2) of
corank more than 3. Let uvw be a face of G. Then there exists a nonzero vector x E ker(M) with x, = x = x,,, = 0. We can assume that x has minimal support. By Van der Hoist's lemma (Theorem 10.1), GI supp+(x) and GI supp-(x) are nonempty and connected. As G is maximally planar, G is 3-connected. Hence there exist three vertex-disjoint paths P1, P2, P3 from supp(x) to {u, v, w}. Let Pf, P2, P3 be the parts outside supp(x). Then the first vertices of the P, belong to N(supp(x)), and hence (by (57)) to both N(supp+(x)) and N(supp-(x)). Contracting each of supp+(x), supp-(x), Pf, P2, P3 to one point, would give an embedded outerplanar graph with uvw forming a face and u, v, and w having two common neighbours. This is not possible.
13
Characterizing A(G) < 3
We characterize in this section the graphs G satisfying A(G) < 3, a result of [12]. The main ingredient is:
Theorem 13.1 If G is planar then A(G) < 3. Proof Suppose G = (V, E) is a planar graph with A(G) > 4 and I V I minimal. We assume that we have an embedding of G in the sphere. For each face f of G let Vf be the set of vertices incident with f. Note that G is 4-connected, since otherwise it would be a subgraph of clique sums of smaller planar graphs, and hence we would have A (G) < 3 by Theorem 6.2. Let ¢: V -+ 1R4 be representative for G. Then q(v) # 0 for each v E V, since otherwise we can delete v, contradicting the minimality of G. So we can assume that 110(v)ll = 1 for each v E V. We may assume that, for each edge uv, 0(u) # ±O(v), since otherwise,
either 0(u) = 0(v), in which case we can contract the edge {u, v} in G, or q5(u) = -q(v), in which case we can delete the edge {u, v} from G. In either case we obtain a contradiction with the minimality of G.
Observe that if f and f are faces with dim(q(Vf)) = dim(O(Vfi)) = 2 and having a common edge, e say, then lin.hull(¢(Vf)) = lin.hull(O(Vp)), as it is equal to lin.hull(q5(e)). Similarly, lin.hull(O(Vf)) C_ lin.hull(¢(Vfi)) if dim(O(Vf)) = 2, dim(q5(Vp)) = 3 and f, f' share a common edge. Fixing V, we choose E maximal under the condition that 0(u) # ±O(v) for each edge {u, v}. Then dim(q5(Vf)) E 12,31 for each face f. Indeed,
dim(O(Vf)) > 2, as each edge e = uv has dim(q5({u, v})) > 2. Moreover, if dim(O(Vf)) = 4, then Vf contains at least two nonadjacent vertices u, v with dim(q({u, v})) = 2. As we can add the edge uv, this contradicts the maximality of E.
Minor-monotone Graph Invariants For c E ii
183
let c+ :_ {v E V I cT¢(v) > 0}, c :_ {v E V I cTO(v) < 0},
( 58)
and let F, be the set of faces f for which Vf intersects both c+ and c-. Then:
Let f and f' be two faces with dim(¢(Vf U Vf,)) = 4. Then
(59)
there is a c E R4 with f, f' E -F,.
To see this, we note that because dim(O(Vf)) > 2, dim(O(Vf,)) > 2, and dim(O(Vf U Vf,)) = 4, there exist vertices u, v E Vf and u', v' E Vp with dim( ({u, v, u', v'})) = 4. Therefore, we can find a c E 1R4 such that u, u' E c+
and v, v' E c. So f, f' E F,, proving (59). For c E R4, let WW := U{VfI If E JFc}. To finish the proof of the theorem, it suffices to show:
dim(O(Wc)) < 3 for each c E R.
(60)
This is sufficient, since (60) implies an immediate contradiction with (59), as there exist faces f and f' with dim(Vf U Vp) = 4, since dim(O(V)) = 4 and
as there is a face f with dim(O(Vf)) = 3 (since if dim(q(Vf)) = 2 for each face f then dim(O(V)) = 2, since lin.hull(¢(Vf)) = lin.hull(q5(Vp)) for any two adjacent faces f, f'). We show that (60) holds. It suffices to show the result for those c with We inclusionwise maximal, and hence with cT cb(v) # 0 for each vertex v. Let such a c be given. As both GIc+ and GIc are connected, the cut 8(c+) corresponds in the dual graph of G to a circuit C which traverses exactly two edges in each face f E Fc. Suppose, to obtain a contradiction, that dim(O(WW)) = 4. Then there
exist faces f, f' E F'c with dim(O(Vf)) = dim(O(Vp)) = 3 and such that lin.hull(¢(Vf)) # lin.hull(q5(Vp)) (as otherwise lin.hull(q5(Vf)) = lin.hull(O(Vf,)) for all f, f' E F, with dim(q5(Vf)) = 3 and dim(¢(Vf,)) = 3, which implies that dim(O(WW)) = 3). They correspond to two vertices on C. Denote by fl, ... , ft the faces between f and f' when travelling from f to f' along C (in a given dir-
ection). Set fo := f and ft := f'. Then we may assume that dim(O(Vf;)) = 2 for all i = 1, ... , t. (Otherwise we can make t smaller.) For i = 0, 1, ... , t, let uivi be the edge common to the faces fi and fi+i So each uivi belongs to 8(c+) (as G is 4-connected). We may assume that ui E c+
and vi E c for each i.
Now choose w E Vf so that ¢(w) ¢ lin.hull(¢(Vp)) and w' E Vp so that t(w') ¢ lin.hull(¢(Vf)). Then the set ¢({uo, vo, w, w'}) has dimension 4. Hence, there exists a d E 1[84 such that cho(w) > 0, &'¢(w') > 0, &' (uo) = 0, and & (vo) = 0. Then the set d+ U d- contains none of the vertices on the faces fl, . . . , f t (since V f C lin.hull(qS({uo, vo})) for all i = 1, ... , t). In particular, ui, vi ¢ d+ U d- for i = 1, . . . , t. By the connectivity of GId+ there exists a path P from w to w' which is entirely contained in d+. 1
Alexander Schrijver
184
t+1O -.fi (where Consider the region R fi is the topological closure of fi) As P joins two vertices on the boundary of R, R U P partitions the rest of the sphere into two regions R1 and R2. We choose indices such that R1 has the vertices uo, ... , ut on its boundary, while R2 has the vertices vo, . . . , vt on its boundary. By the connectivity of G1d-, d- is contained either in R1 or in R2. Suppose Ui
first that d- is contained in R1. Consider the vector d = d + Ec, with E > 0 _ d+ and d- C d-. Then, d- D {vo, ... , vt} U d-, small enough such that d+ C . . , ut E d+. Then there is no path joining vo and d- which is entirely contained in d-, contradicting the connectivity of G1d-. If d- is contained in R2, we arrive similarly at a contradiction, by consider-
while uo, .
ingd=d - ec. We can now characterize the graphs G satisfying A(G) < 3. Having Theorem 13.1, Theorem 6.2 gives that A(G) < 3 also holds for graphs G obtained from planar graphs by taking clique sums and subgraphs. This characterizes the graphs G with A (G) < 3, as follows from the following two theorems.
Theorem 13.2 If G has no K5- or V8-minor, then G can be obtained by taking clique sums and subgraphs from planar graphs.
Proof Suppose G is not planar. If G is not 3-connected, then it is easy to see that G is a subgraph of a clique sum of two smaller graphs not having a K5or V8-minor. So we may assume that G is 3-connected. Then by Wagner's theorem [20], G can be obtained as a subgraph of a 3-clique sum of two smaller graphs Gl and G2 both with no K5-minor. Let K be the clique. It suffices to show that Gl and G2 have no V8-minor. Suppose to the contrary that G1, say, has a V8-minor. As V8 does not contain any triangle, the V8-minor in Gl does not need all three edges of K. So Gl - e has a V8-minor for some edge e in K. However, Gl - e is a minor of G (by the 3-connectedness of G), contradicting the fact that G does not have a V8-minor.
In [12] also the following was shown (we thank Andries Brouwer for communicating the proof below to us):
Theorem 13.3 A(V8) = 4. Proof The inequality \(V8) < 4 follows from the fact that for any vertex v of V8, the graph V8 - v is planar. Hence A(V8) < )(V8 - v) + 1 < 4 by Theorem 13.1.
Minor-monotone Graph Invariants
185
We next show A(V8) > 4. Represent V8 as the graph G with vertex set V = {0, . . . , 7}. Let M = (mi,j) be the 8 x 8 matrix defined by 1
mi,j
=
ifi=j,
-v ifJi -jJ =1 or 7, -1
ifJi -jl = 4,
=0
otherwise,
(61)
where we assume that the rows and columns are labelled 0, ... , 7.
One can show that M has rank at most 4 as follows. For a = 17ri and a = 17ri, let x' be the vector in C8 defined by := ear
for j = 0,
(62)
These two vectors are linearly independent and both satisfy Mx" = 0. Indeed, by symmetry it suffices to show that for both choices of a one has (Mxa)o = 0. Now . . .
, 7.
(Mxa)o=xo -x'-lV"2-xi
-x4 =1-e-'v/2-e*v2-e4a
(63)
taking subscripts mod 8. If a = 47ri, then e' + e-a = 2 cos 47r = , while e4a = cos 7r + i sin 7r = -1, and hence (Mxa )o = 0. If a = a i7r, then ea + e-a = 2 cos 2 = 0, while e4a = cos 2ir + i sin 27r = 1, and hence again (Mxa)o = 0. Since the real and imaginary parts of the two vectors xa give four vectors linearly independent over R, we know that corank(M) > 4. Let X be the kernel (null space) of M. We show that X is representative for G. Choose a nonzero x E X. So
xj =xj-1V2+x,j+lV 2+xj+4.
(64)
Let W := supp+(x). Then W # 0, since otherwise for any j with xj+l < 0, the value of xj would be strictly smaller than xj+l by (64). Assume that W induces a disconnected subgraph of V8. Let U := V \ W, and let K1 and K2 be two of the components of GIW. Then IKil > 2, since otherwise Ki would consist of one vertex, contradicting (64). So JUG < 4. Since V8 is 3-connected, since each cutset of size 3 consists of the set of vertices
adjacent to one vertex, and since U separates K1 and K2, it follows that IUD = 4, and that the subgraph induced by W consists of two disjoint edges. Now for each edge e = {j, j + 1} of V8, each other edge e' of V8 disjoint from e contains at least one vertex that is adjacent to at least one vertex in e.
It follows that W = {1, 3, 5, 7} or W = {0, 2, 4, 6}. Then (64) implies that xj < xj+4 for each j E W, and hence xj = xj+4 for each j E W. But then xj = 0 for each j E U, contradicting the fact that supp- (x) 0. Thus we have the following theorem:
Alexander Schrijver
186
Theorem 13.4 Let G be a graph. Then )t(G) < 3 if and only if G arises by taking clique sums and subgraphs from planar graphs; that is, if and only if G has no K5- or V8-minor.
Proof Directly from Theorems 5.2, 6.2, 13.1, 13.2, and 13.3.
14
A Borsuk theorem for antipodal links
We next come to studying µ(G) < 4 and )(G) < 4. The following Borsuktype theorem on the existence of certain antipodal links is essential in the proof. This theorem, for general dimension, is proved in [15].
Let P be a convex polytope in R. We say that two faces F and F' are antipodal if there exists a nonzero vector c in 1[8" such that the linear function cT x is maximized by every point of F and minimized by every point of F'. Let (P)1 denote the 1-skeleton of P. For any face F of P, let aF be its boundary.
Theorem 14.1 Let P be a full-dimensional convex polytope in 1[85 and let 0 be an embedding of (P)1 into 1183. Then there exists a pair of antipodal 2-faces F and F' such that O(aF) and q5(aF') are linked. In [15] this is derived (for general dimension) from a Borsuk-type theorem on the existence of antipodal intersections, extending a result of Bajmoczy and Barany [2] slightly. A direct proof of Theorem 14.1 can be sketched as follows. First:
We can assume that if F and F' are antipodal 2-faces of P, then F - F and F' - F' do not have any nonzero vector in
(65)
common.
This can be shown by applying a small projective perturbation to P. For any two disjoint closed curves C and C' in R3, let lk(C, C) denote their linking number, which is the number mod 2 of crossings in any link diagram where C is over C. (This is a topological invariant.) Then:
There exists an embedding b: (P)1 -+ 1[83 with the property
(66)
that there is exactly one pair of antipodal 2-faces F, F' for which lk(O(aF),O(aF')) = 1. Indeed, we can assume that by maximizing the last coordinate x5 we obtain some 2-face FO and by minimizing x5 we obtain some 2-face FO' antipodal to FO. Moreover, we can assume that (0, 0, 0, 0, 1) belongs to the relative interior of Fo
and that (0, 0, 0, 0, -1) belongs to the relative interior of F0'. For any vector x = (x1,. .. , x5) in 1R5, let i := (x1i ... , x4). Then define z&: (P)1 -4 S3 by: O(x)
x
(67)
Minor-monotone Graph Invariants
187
We can assume that z/i is an embedding of (P)1 into S3 (by moving P slightly).
Then Fo, FO' is the only pair of antipodal 2-faces F, F' for which we have lk(O(OF), O(OF')) = 1. Finally:
We can deform 0 to 0 while only edges are moved
(68)
through each other; at each such operation, the quantity E lk(aF, aF') remains invariant. F,F'
This follows from the fact that for any two edges e, e' of (P)1, the number of pairs of antipodal 2-faces F, F' with e C OF and e' C OF', is even. 15
Characterizing p(G) < 4
We first give a brief introduction to the work of Robertson, Seymour, and Thomas on linklessly embeddable graphs. An embedding of a graph G into R3 is called linkless if any two disjoint circuits in G have unlinked images in R. A graph G is linklessly embeddable (in R3) if it has a linkless embedding in R. There are a number of equivalent characterizations of linklessly embeddable graphs. Call an embedding of G flat if for each circuit C in G there is a
disk D (a `panel') disjoint from (the embedding of) G and having boundary equal to C. Clearly, each flat embedding is linkless, but the reverse does not hold. (For instance, if G is just a circuit C, then any embedding of G is linkless, but only the unknotted embeddings are flat.) However, if G has a linkless embedding, it also has a flat embedding. So the collections of linklessly
embeddable graphs and of flatly embeddable graphs are the same. This was shown by Robertson, Seymour, and Thomas [19], as a byproduct of a proof of an even deeper forbidden-minor characterization of linklessly embeddable graphs. To understand this forbidden-minor characterization, it is important to note that the class of linklessly embeddable graphs is closed under the YAand AY-operations. It implies that also the class of forbidden minors for linkless embeddability is closed under applying YA and AY. Now Robertson, Seymour, and Thomas [19] showed:
the Petersen family is the collection of forbidden minors for
(69)
linkless embeddability.
Here the Petersen family is the class of graphs arising from the Petersen graph by any series of AY- and YO-operations. The Petersen family consists of seven graphs, and includes the graph K6 (see Figure 2). It turns out not to be difficult to prove that p(G) = 5 for each graph in the Petersen family. In fact, by result (51) of Bacher and Colin de Verdiere [1], the class of graphs G with µ(G) = 5 is closed under AY and YA. Since moreover µ(K6) = 5, we know µ(G) = 5 for each graph G in the Petersen family.
Alexander Schrijver
188
Figure 2: The Petersen family So with the forbidden minor characterization of Robertson, Seymour, and Thomas [19], we know that if p(G) < 4 then G is linklessly embeddable. The reverse implication was conjectured by Robertson, Seymour, and Thomas [17] and proved in [15], and thus we have:
Theorem 15.1 µ(G) < 4 if and only if G is linklessly embeddable; that is, if and only if G does not have a minor in the Petersen family. We will not give the full proof of this here, but rather give an indication of the proof by showing that A(G) < 4 for each linklessly embeddable graph G. The proof that also p(G) < 4 for linklessly embeddable graphs G is similar, but requires a few more technicalities, and we do not give it in this paper.
Minor-monotone Graph Invariants
16
189
Towards characterizing A(G) < 4
We do not know a complete characterization of the class of graphs G satisfying A(G) < 4. However, we have ([15]):
Theorem 16.1 If G is linklessly embeddable, then A(G) < 4. Proof Let G be linklessly embedded in W, and suppose that A(G) > 5. Then there is a 5-dimensional subspace L of 18V such that Gl supp+(x) is nonempty and connected for each nonzero x E L. Call two elements x and x' of L equivalent if supp+(x) = supp+(x') and supp- (x) = supp- (x'). The equivalence classes decompose L into a centrally symmetric complex P of pointed polyhedral cones. Choose a sufficiently dense set of vectors of unit length from every cone in P, in a centrally symmetric fash-
ion, and let P be the convex hull of these vectors. Then P is a 5-dimensional centrally symmetric convex polytope such that every face of P is contained in a cone of P.
We define an embedding 0 of (P)1 in W. For each vertex v of P, we choose a vertex v' of G in supp+(v), and we let O(x) be a point in R3 very near v'. For each edge e = uv of P, we choose a path e' connecting u' and v' in GI supp+(x), where x is an interior point of e. (By our construction, supp+(x) is independent of the choice of x, and contains both supp+(u) and supp+(v).) Then we map e onto a Jordan curve connecting O(u) and O(v) very near e'. Clearly we can choose the images of the vertices and edges so that this map ¢ is one-to-one. Then by Theorem 14.1, P has two antipodal 2-faces F and F' such that the images of their boundaries are linked. Since P is centrally symmetric, there is a facet D of P such that F C D and F' C -D. Let y be a vector in the interior
of D. Then the images of OF and aF' are very near subgraphs spanned by supp+(y) and supp-(y), respectively, and hence some circuit of G spanned by supp+(y) must be linked with some circuit in supp-(y), a contradiction.
Corollary 16.2 If G is obtained from linklessly embedded graphs by taking clique sums and subgraphs, then a(G) < 4. Proof Directly from Theorems 6.2 and 16.1.
By Theorem 5.2, G = K6 is a forbidden minor for the class of graphs G with A(G) < 4. Any other graph G in the Petersen family of graphs however satisfies A(G) < 4, since:
Theorem 16.3 Let G be in the Petersen family with G # K6. Then G is obtainable by taking clique sums and subgraphs from K5.
Alexander Schrijver
190
Proof Inspection of the Petersen family (Figure 2) shows that G is either a subgraph of the graph obtained from K7 by deleting the edges of a triangle, and this graph is a clique sum of three K5's, or G arises from such a subgraph by one or more AY-transformations, that is, it is a subgraph of a clique sum with K4's.
This immediately implies that A(G) < 4 for each graph G K6 in the Petersen family. Moreover, it follows that each such graph is obtainable by taking clique sums and subgraphs from linklessly embeddable graphs. Note that the graph G obtained from V8 by adding a new vertex adjacent to all vertices of V8, cannot be obtained from linklessly embeddable graphs by taking clique sums and subgraphs; but G does not have a K6-minor. In fact, it satisfies A(G) = 5. However it is not minor-minimal for the property
)(G) > 5. Let V9' arise from V8 by adding an extra vertex vo, adjacent to v2, v4, v6, v7, v8 (see Figure 3). Similarly, let V9" arise from V8 by adding an extra vertex vo
Figure 3: The graph V9
Figure 4: The graph V9"
adjacent to v2i v3, v5, v7, v8 (see Figure 4). It is shown in [12] that V9 and V9" are minor-minimal graphs G with A(G) > 5. The graphs V9 and V9' are also minor-minimal graphs not obtainable from linklessly embeddable graphs by taking clique sums and subgraphs. This can be seen as follows. Since A(V9) = )(V9') = 5, it follows from Corollary 16.2 that these two graphs indeed are not obtainable in such a way. Moreover,
to see that they are minor-minimal, observe that deleting or contracting any edge of V9 or V9', produces a graph that has a vertex whose deletion makes the graph a clique sum of planar graphs. Since the class of graphs G with A(G) < 4 is closed under taking AY operations (not under YD), we can obtain other graphs with A(G) > 5 by applying a YA operation to V9 or V9". Any of them contains a K6-minor, except if we apply YA to vertex vl (or equivalently, to v5) of V.
Minor-monotone Graph Invariants
17
191
An extension to oriented matroids
It turns out that the results described above for A(G) can be extended to oriented matroids, as is shown in [7]. Before describing this, we first give the definition of and a little further background on oriented matroids (see Bjorner, Las Vergnas, Sturmfels, White, and Ziegler [3] for more information). It is convenient to introduce, for any ordered pair x = (a, b), the notation
x+ := a and x- := b, and -x = (b, a). An oriented matroid (V, X) consists of a finite set V and a collection X of ordered pairs x = (x+, x-) of subsets of V such that:
(i) for each x E X, x+ fl x- = 0;
(70)
(ii) 0 := (0, 0) E X;
(iii) if X EX then -x E X;
(iv) if x, y E X, then x y E X where x y is defined by
x-y:=(x+U(y+\x ),x U(y \x+)); (v) if x, y E X and U E x+ fly , then there exists a z E X such that u z+Uz-, (x+\y)U(y+\x-) C z+ C_ x+Uy+,
and (x-\y+)U(y\x+) Cz- Cx- Uy. The elements of X are called the vectors of the oriented matroid. (0 is the zero.) Any linear subspace Y of 1W gives an oriented matroid (V, X), by taking
X := {(supp+(x), supp-(x)) I x E Y}.
(71)
For any oriented matroid M = (V, X ), the minimal nonempty subsets of {x+ U X- I x E X} form the circuit collection of a matroid, again denoted
by M. Thus matroid terminology applies to oriented matroids, and we can speak of the rank rank(M) of an oriented matroid M: it is the maximum size of a subset of V not containing any circuit as a subset. The corank corank(M) of M is equal to IV) - rank(M). It is not difficult to prove that if M is given by (71), then corank(M) = dim(Y). (72)
Now the graph parameter A'(G) is defined as follows. Let G = (V, E) be an undirected graph. An oriented matroid M = (V, X) is called representative for G if
for each nonzero x E X, x+ is nonempty and induces a connected subgraph of G.
(73)
Then A'(G) is the largest corank of an oriented matroid representative for G. From (72) one derives that for each graph G: A(G) < A'(G).
(74)
Alexander Schrijver
192
One of the consequences of the results described below is that there are no graphs G with A(G) < 3 and A(G) < A'(G). In fact, we do not know any graph G with strict inequality in (74). Any result we know for A(G), also holds for A'(G). First of all, A'(G) is minor-monotone:
if G is a minor of H then )'(G) < A'(H).
(75)
Moreover one has:
A'(K,,) = n - 1.
(76)
So again Hadwiger's conjecture implies the conjecture that y(G) < A'(G) + 1 for each graph G, where -y(G) is the colouring number of G. Moreover:
For any graph G and vertex v of G one has A'(G - v) >
(77)
A'(G) - 1. Again for each t > 1 the class of graphs G with A(G) < t is closed under taking clique sums, since:
If G has at least one edge and is a clique sum of Gl and G2,
(78)
then A'(G) = max{A'(G1), A'(G2)}.
This directly implies characterizations of those graphs G satisfying A'(G) < 1 and A'(G) < 2:
A'(G) < 1 if and only if G is a forest,
(79)
A'(G) < 2 if and only if G is a series-parallel graph.
(80)
and
Moreover, it can be proved that a graph G satisfies A'(G) < 3 if and only if G can be obtained from planar graphs by taking clique sums and subgraphs.
(81)
Recently, Rudi Pendavingh showed: if G is obtainable from linklessly embeddable graphs by taking subgraphs and clique sums, then A'(G) < 4.
18
(82)
The related graph invariant k(G)
We finally describe a graph invariant related to A(G) (introduced in [12]), for which the set of forbidden minors can be precisely characterized. For any
Minor-monotone Graph Invariants
193
connected graph G = (V, E), define ic(G) to be the largest d for which there exists a function 0 : V -* Rd such that: (i) O(V) affinely spans a d-dimensional affine space;
(83)
(ii) for each affine halfspace H of Rd, ¢-'(H) induces a connected subgraph of G (possibly empty).
(An affine halfspace is a set of the form {x
cTx > 8} for some nonzero vector c.) Note that such a function 0 does not exist for disconnected graphs; so c(G) is undefined if G is disconnected. Observe that if G is the 1-skeleton of a full-dimensional polytope in Rd, then ic(G) > d, as the polytope gives the embedding in 1[8d. By similar arguments as used in the proof of Theorem 3.1 one shows that if H is a connected minor of G then c(H) < ic(G). So again for each d there is a finite collection of forbidden minors for the collection of graphs satisfying ic(G) < d. This collection of graphs is equal to {Kd+2}, as is shown in the next theorem. First observe that ic(G) < A(G)
I
(84)
holds for each connected graph G, since if 0: V -* 1[8d satisfies (83), then we may assume that the origin belongs to the interior of the convex hull of O(V). But then trivially ¢ is representative for G. Basic in the characterization is the following observation (Grunbaum and Motzkin [9], Grunbaum [8]):
Theorem 18.1 If G is the 1-skeleton of a d-dimensional polytope P, then G has a Kd+l-minor.
Proof By induction on d, the case d = 0 being trivial. If d > 0, let F be a facet of P. By the induction hypothesis, the 1-skeleton of F can be contracted
to Kd. Moreover, the vertices of P not on F induce a connected subgraph of G, and hence can be contracted to one vertex. This yields a contraction of G to Kd+l, as each vertex of F is adjacent to at least one vertex of P not on F. This gives:
Theorem 18.2 For each connected graph G and each d, s'c(G) > d if and only if G has a Kd+l-minor. Proof Sufficiency. One has ic(Kd+1) = d since the vertices of a simplex in Rd give a function 0 satisfying (83). So if G has a Kd+l-minor, then ,c(G) > d.
Alexander Schrijver
194
Necessity. Let G = (V, E) be a connected graph and let d := ic(G), such that for each proper connected minor H one has ic(H) < d. By Theorem 18.1 it suffices to show that G is the 1-skeleton of a d-dimensional polytope. Let 0: V -* Rd satisfy (83). Let P be the convex hull of O(V). So P is a d-dimensional polytope in Rd. We show that G is the 1-skeleton of P. First observe that for each vertex x of P, the set 0-1 (x) induces a connected subgraph of G, as it is equal to 0-1(H) for some affine halfspace H of Rd. Hence if ¢-1(x) consists of more than one vertex of G, then we can contract this subgraph to one vertex, contradicting the minimality of G. Similarly, for each edge xy of P, the set 0-1(xy) induces a connected subgraph of G. Hence it contains a path from ¢-1(x) to ¢-1(y). As this is true for each edge, G contains a subdivision of the 1-skeleton of P as a subgraph. By the minimality of G this implies that G is equal to the 1-skeleton of P. So Hadwiger's conjecture is equivalent to y(G) < c(G) + 1 for each connected graph G.
Acknowledgements We thank the referee for several suggestions improving the presentation.
References [1] R. Bacher & Y. Colin de Verdiere, Multiplicites des valeurs propres et transformations etoile-triangle des graphes, Bulletin de la Society Mathymatique de France, 123 (1995), 101-117. [2] E. G. Bajmdczy & I. Barany, On a common generalization of Borsuk's and Radon's theorem, Acta Mathematica Academiae Scientiarum Hungaricae, 34 (1979), 347-350. [3] A. Bjorner, M. Las Vergnas, B. Sturmfels, N. White & G. Ziegler, Oriented Matroids, Cambridge University Press, Cambridge (1993).
[4] S. Y. Cheng, Eigenfunctions and nodal sets, Commentarii Mathematici Helvetici, 51 (1976), 43-55. [5] Y. Colin de Verdiere, Sur un nouvel invariant des graphes et un critere de planarite, Journal of Combinatorial Theory, Series B, 50 (1990), 11-21.
[6] Y. Colin de Verdiere, On a new graph invariant and a criterion for planarity, in Graph Structure Theory (Proceedings of the AMS-IMS-SIAM Joint Summer Research Conference on Graph Minors, Seattle, 1991) (eds. N. Robertson & P. Seymour), Contemporary Mathematics, 147, American Mathematical Society, Providence, Rhode Island (1993), 137-147.
Minor-monotone Graph Invariants
195
[7] J. Edmonds, M. Laurent & A. Schrijver, A minor-monotone graph parameter based on oriented matroids, Discrete Mathematics, in press. [8] B. Grunbaum, On the facial structure of convex polytopes, Bulletin of the American Mathematical Society, 71 (1965), 559-560.
[9] B. Griinbaum & T S. Motzkin, On polyhedral graphs, in Convexity (ed. V. Klee), Proceedings of Symposia in Pure Mathematics, 7, American Mathematical Society, Providence, Rhode Island (1963), pp. 285-290. [10] H. van der Holst, A short proof of the planarity characterization of Colin de Verdiere, Journal of Combinatorial Theory, Series B, 65 (1995), 269272.
[11] H. van der Holst, Topological and Spectral Graph Characterizations, Ph.D. Thesis, University of Amsterdam, 1996. [12] H. van der Holst, M. Laurent & A. Schrijver, On a minor-monotone graph invariant, Journal of Combinatorial Theory, Series B, 65 (1995), 291-304.
[13] H. van der Holst, L. Lovasz & A. Schrijver, On the invariance of Colin de Verdiere's graph parameter under clique sums, Linear Algebra and its Applications, 226 (1995), 509-517. [14] A. Kotlov, L. Lovasz & S. Vempala, The Colin de Verdiere number and sphere representations of a graph, preprint, 1996.
[15] L. Lovasz & A. Schrijver, A Borsuk theorem for antipodal links and a spectral characterization of linklessly embeddable graphs, Proceedings of the American Mathematical Society, in press. [16] N. Robertson & P. D. Seymour, Graph minors. XX. Wagner's conjecture, preprint, 1988.
[17] N. Robertson, P. D. Seymour & R. Thomas, A survey of linkless embeddings, in Graph Structure Theory (Proceedings of the AMS-IMS-SIAM Joint Summer Research Conference on Graph Minors, Seattle, 1991) (eds. N. Robertson & P. Seymour), Contemporary Mathematics, 147, American Mathematical Society, Providence, Rhode Island (1993), pp. 125-136.
[18] N. Robertson, P. Seymour & R. Thomas, Hadwiger's conjecture for K6free graphs, Combinatorica, 13 (1993), 279-361. [19] N. Robertson, P. Seymour & R. Thomas, Sachs' linkless embedding conjecture, Journal of Combinatorial Theory, Series B, 64 (1995), 185-227.
[20] K. Wagner, Uber eine Eigenschaft der ebene Komplexe, Mathematische Annalen, 114 (1937), 570-590.
196
Alexander Schrijver CWI Kruislaan 413 1098 SJ Amsterdam The Netherlands and Department of Mathematics University of Amsterdam Plantage Muidergracht 24 1018 TV Amsterdam The Netherlands
[email protected] Some Applications of Algebraic Curves in Finite Geometry and Combinatorics T. Szonyi Summary Various applications of Weil's theorem in finite geometry and combinatorics are surveyed. Several illustrative proofs are sketched. As a by-product, we give an up-to-date account on what is known about complete arcs, minimal blocking sets and (k, n)-arcs in the desarguesian plane PG(2, q). 1
Introduction
Four years ago at the British Combinatorial Conference, Aart Blokhuis [16] gave a lecture on some applications of polynomials in finite geometry and combinatorics. The aim of this survey is to collect those applications of polynomials where the polynomial is considered as a curve, or more generally as an algebraic variety over a finite field. Most of the applications depend on estimates for the number of GF(q)-rational points of curves. In some of the applications Bezout's theorem is enough, but typically the deep theorem of Weil, or a variant or consequence of it, is used. Weil's theorem gives a very strong bound on the number of GF(q)-rational points of a curve: Let f be an absolutely irreducible curve of degree d, defined over GF(q), and denote by N the number of GF(q)-rational points on it. Then
q-
+1 1. If k > 3t and q >
4t4k21-2
then the
system
F _ {0(A) : A C V1i IAA =t-11 is self-avoiding.
In a previous version of their paper, Babai, Gal and Wigderson [4] used different bipartite graphs to construct large self-avoiding Sperner families, based on a recent work by Kollar, R6nyai and Szab6 [72] concerning Zarankiewicz' problem. Since the key ingredient in [72] is to estimate the number of solutions of a system of equations, we include their result here.
Let H be a fixed graph. The Turan number ex(n, H) is the maximum number of edges in a graph on n vertices which does not contain a copy of H. Zarankiewicz' problem is to determine the Turin number for (complete) bipartite graphs. Let Kt,, denote the complete bipartite graph on t + s vertices
and with is edges. For a fixed t and s > t, Kovari, T. S6s and Turin [74] proved that ex(n, Kt,s) < ct,9n r,
where ct,s is a constant depending on s and t. The best general lower bound, obtained by the probabilistic method, yields only cn
2- a+1-2
st-1 < ex (n, Kts),
where c' is a positive absolute constant. Using the so-called norm-graph, Kollar, R6nyai and Szab6 showed that, for s > t! + 1 the Kovari, T. S6s, Turin bound gives the right order of magnitude ex (n, Kt,3) > 2-tn2
It
.
(2)
Their construction is algebraic. The norm-graph Gq,t = G is defined as follows: the set of vertices of G is GF(gt). Recall that the norm (with respect to GF(q))
of an element a E GF(qt) is N(a) = a aq ..... aqt-1 = a(q`-1)/(q-1). It is wellknown (see e.g. [76]) that N(x) E GF(q) and for any non-zero u E GF(q), the
206
T. Szonyi
number of solutions of N(x) = u is (qt - 1)/(q - 1). Two vertices a, b of G are connected by an edge if N(a + b) = 1. For this graph Kollar, Ronyai and Szabb proved the following.
Theorem 3.6 (Kollar, Ronyai, Szabo [72]) The graph G = Gq,t contains no subgraph isomorphic to Kt,t!+l Let us sketch the proof: if dl, ... , dt are t different elements of G, then the vertices adjacent to all of them are the solutions of the system of equations N(x + di) = 1, for i = 1, ... , t. Using N(x + di) = (x + di) (x° + di°) x9`-1
(xqt-1
+ dqt-1)
di_19'-1
replacing by xi and by -aid and putting arbitrary bi's on the right-hand side one gets the following system of equations: bl
- alt)
= =
- att)
=
bt.
(x1 - all)(x2 -
a12)...(xt
- alt)
(x1 - a21)(x2 -
a22)...(xt
(x1 - atl) (x2 -
at2)...(xt
b2
The key step in the proof of the theorem is to show that such a system of equations has at most t! solutions if the bi's are different. Note that the highest degree terms on the left-hand side are the same in each equation, so the equations as hypersurfaces meet in a variety of codimension 2 (contained in the hyperplane at infinity). The result says that their affine parts intersect in a finite number (at most t!) of points. The previous results showed that among the values of a polynomial, that is not a constant multiple of a square of another polynomial, the squares and non-squares are distributed almost evenly. The classical inequality of Polya and Vinogradov says that the same property holds for intervals in prime fields. It gives that for any h h
EW(i)
J loge.
(3)
i=0
This can be used to study another interesting class of algebraically defined
graphs: the graphs B(n, t). These graphs are also (p, a)-jumbled, see [110]. The vertices of B(n, t) are the elements of GF(n), where n is prime. Two vertices x, y are joined by an edge if and only if (x - y)2 E {1, ... , t}. These graphs are regular with degree d = I {x : x2 E {1,. .. , t}}J. Using the PolyaVinogradov inequality one sees that d/t --+ 1, if t > n1/4 log n. Bollobas [28] has found a pleasing proof, based on the Polya-Vinogradov inequality, of the following result.
Theorem 3.7 No two vertices of the graph B(n, t) have more than t2/n + log2 n common neighbours.
Algebraic Curves in Finite Geometry
207
Further properties of B(n, t) can be found in Thomason [110], together with references to improvements of the P61ya-Vinogradov inequality. Applications of the Polya-Vinogradov inequality in finite geometry, and a far reaching generalization due to Smith, can be found in Section 4.
4
Applications of Weil's theorem in finite geometry Let us define some important objects in finite geometry.
Definition A (k, n) -arc in a projective plane of order q is a set of k points with some n but no n + 1 points on a line. (k, 2)-arcs are simply called k-arcs.
A k-arc is complete if it is not contained in a (k + 1)-arc, that is when it is maximal subject to inclusion. Similarly a (k, n)-arc is complete if it is not contained in a (k + 1, n)-arc. Definition An s-fold blocking set K in PG(2, q) (or AG(2, q)) is a set of points such that every line of PG(2, q) (or AG(2, q)) intersects K in at least s points. A 1-fold blocking set is simply called a blocking set. We say that a (1-fold) blocking set is trivial if it contains a line of PG(2, q). An s-fold blocking set is called minimal or irreducible when no proper subset of it is an s-fold blocking set. For s = 2 and 3 we also speak of a double blocking set and a triple blocking set.
Note that (k, n)-arcs and s-fold blocking sets with n + s = q + 1 are in fact each other's complement. Typically, the (k, n)-arc terminology is used if n is small compared to q and we use the s-fold blocking set terminology ifs is small. The aim of this section is to collect some results, mainly constructions, for these objects. Several good survey papers are recommended: Hirschfeld [59] and Hirschfeld-Storme [64] are general ones, whereas Blokhuis [19] is mainly about blocking sets.
Throughout this paper we use the usual representation of AG(2, q) and PG(2, q). This means that the points of AG(2, q) have affine coordinates (x, y) where x, y are elements of GF(q). The lines of this affine plane have equation mX + b - Y = 0 or X = c. The coefficient m is the slope of the line, and the infinite points of PG(2, q) can be identified with slopes. So (m) will denote the infinite point of lines with slope m. Similarly (oo) will be the infinite point of vertical lines, that is lines with equation X = c. We shall also use the following standard terminology: if K is a set of points
and 2 is a line intersecting K in exactly s points, then we call 2 an s-secant of K. Instead of 1-secant also the expression tangent will be used. The first natural question for k-arcs is to determine the maximum value of k. This was done by Bose (see Chapter 9 of [58]), who proved that k < q+2 and showed that k = q + 2 is only possible if q is even. (q + 1)-arcs are called
208
T. Sz6nyi
ovals, (q + 2)-arcs are called hyperovals. There are natural examples: a conic is always a (q + 1)-arc. For q even the tangents of a conic pass through a point, called the nucleus of the conic. Adding this point results in a hyperoval. The next step is to go beyond Bose's bound. If a k-arc consists of slightly fewer than q + 1 or q + 2 points, then it can always be embedded in an arc having the maximum number of points. This beautiful theory is due to Beniamino Segre [90], and we discuss his results and some recent achievements in the next section. Note also that for q odd, Segre proved that (q + 1)-arcs are in fact conics, so very large arcs are completely described. Let us introduce a notation: m'(2, q) denotes the size of the largest complete arc, which is not an oval or hyperoval. This implies that a k-arc with k > m'(2, q) can be embedded in a (hyper)oval. From the other end of the spectrum, Lunelli and Sce [79] proved that a complete k-arc has to have at least k > 2q points. Here the important question is to get close to this theoretical lower bound. The size of the smallest complete arc in PG(2, q) will be denoted by n(2, q). Most of the early constructions of complete arcs other than ovals or hyperovals give k-arcs with k > (q + 3)/2, but they all have k - q/2, when completeness is known. For a list of references, see [99]. The first example of a complete k-arc with k - q/3 was given by V. Abatangelo [1]. Before discussing his construction let us describe in general how the method based on Weil's theorem works for proving completeness of arcs. Most of the constructions use the following general suggestion of Segre and Lombardo-Radice [75]: the points of the arc are chosen (with some exceptions) among the points of a conic or a cubic. So let us start with a conic or cubic C. The steps of the general construction scheme are the following:
1. Choose an algebraically parametrized subset K C C. Preferably the parametrization should have low degree.
2. Construct an algebraic curve describing the collinearity of two points of K and a point P outside C.
3. Show that the curve is absolutely irreducible for most P V C. Then Weil's theorem will guarantee the existence of two points of K such that
the line joining them passes through P; in other words, P cannot be added to K. 4. Extend the arc with some exceptional points P
C, and with some
points of C.
Let us illustrate this scheme by a result of V. Abatangelo [1].
1. Start with the parabola C : y = x2. Let the algebraically parametrized subset K be K = {(u3, u6) : 0 # u E GF(q)}, where 3 divides q - 1. Note that IKI = (q - 1)/3.
Algebraic Curves in Finite Geometry
209
2. Take first an affine point P : (a, b) with b # a2. The collinearity of Pl : (x3, x6), P2 : (y3, y6) (PI # p2) and P means that a(x3+y3) _x3y3 _ b = 0. This is the equation of a curve .P of degree 6. Here we used that Pl # P2 implies x3 1 y3.
3. To show that F is absolutely irreducible the following criterion of Segre [89] can be used.
Lemma 4.1 Let F be a plane algebraic curve of degree k over a field E.F is irreducible over the algebraic closure of E if there is a point P on F such that there is no linear component ofF through P and there is a tangent r at P which counts once among the tangents at P and intersects F with multiplicity k. Note that the conditions for the tangent r are easy to verify, since we only have to prove that r f1.P = {P} for the multiplicity condition and the tangents can be found explicitly. In our case, if a 54 0, the conditions can be verified for the infinite points of Y. These are (0, 1, 0) and (1, 0, 0) and the tangents are
the lines x = c and y = d with c3 = d3 = a. (For a = 0 the tangents would not be distinct.) Substituting x = c for such a c in the equation of .P gives a2 - b = 0 and this is not zero. Hence F is indeed absolutely irreducible. The curve is of degree 6 and it has two ordinary singular points with multiplicity 3, so its genus is at most 4 and Weil's theorem shows that there are points with x3 y3 if q is large enough. A similar argument works for infinite points. Doing this precisely, Abatangelo [1] proved the following theorem.
Theorem 4.2 Let q = 2h, h > 8 and K = {(u3, u6) 0 54 u E GF(q)}. If K U {P} is an arc, then either P belongs to the line x = 0 or P is the nucleus (1, 0, 0) of the parabola C or P E C \ K. 4. We can add to K the nucleus (1, 0, 0) and the points (0, g) and (0, g2) of the line x = 0, where g is a generator of the multiplicative group of GF(q).
Abatangelo's idea was generalized to arbitrary multiplicative subgroups by Korchmaros [73].
The same idea was used by the present author [94] for cubic curves, to construct complete k-arcs with k = o(q). To completely follow the above scheme one has to start with a rational curve, see [94, 95]. However, the method can be extended to non-singular cubics but instead of the parametrization, that is a morphism to the projective line, morphisms to the curve itself have to be considered. Let us first mention that for cubic curves one can always work in
an abelian group, since there is an abelian group describing collinearity, see Schoof [86]. In particular, if this abelian group has even order, then the coset of a subgroup of index two will be an arc containing half the points of the cubic. Voloch [115] proved that if the j-invariant of the elliptic cubic is non-zero and
210
T. Szonyi
q > 175 (if q is odd), or q > 256 (if q is even), then these arcs are complete. This construction, without proving the completeness, is due to Zirilli [122]. A nice additional property of these arcs is that they intersect a conic in at most 6 points by Bezout's theorem; for a general combinatorial problem related to this fact, see Cameron [41]. In a second paper Voloch [116] extended his method to cosets of other subgroups. Under some mild conditions, he proved that if the characteristic is not 2 and the j-invariant is not zero, then the coset of a subgroup of index m covers all the points outside the cubic curve provided
that q > 98m4 + 34m2 + 2m(7m2 + 1) 49m2 + 20.
Using essentially this result, the present author [96] proved by a construction that n(2, q) < cq3/4. Very recently Kim and Vu [71] used a sophisticated version of the probabilistic method, the so-called Rodl nibble, to show that n(2, q) < d,,Fq logc q, where d and c are absolute constants. A next natural question is to determine the spectrum of complete arcs, that is to decide which are the possible cardinalities of complete k-arcs. Combining the results of Voloch [115], Hadnagy [55] and Szonyi [96] the following result can be obtained.
Theorem 4.3 (Hadnagy, Voloch, Szonyi) Let p > 510 be a prime number. For every integer k satisfying [2.46 . logp p3/4]
k < (/+ 1)2/2
there exists a complete arc with k points in PG(2,p).
Historically, Voloch proved the part (/ - 1)2 < k < (j + 1)2 by using the complete arcs consisting of half the points of a non-singular cubic curve. He only needed p > 175, and his result extends to planes of non-prime order. Taking points of a conic, Szonyi (see [96]) showed that for every integer k satisfying [p/3] + 3 < k < [p/2] + 1 there exists a complete k-arc in PG (2, p), if p is a sufficiently large prime. Finally, Eva Hadnagy dealt with the longest part of the interval. She needs the bound p > 510, which guarantees for example that the intervals are non-empty. For the proof of completeness, Hadnagy uses a theorem of Smith about small solutions of congruences, which is somewhat similar to the Polya-Vinogradov inequality.
Theorem 4.4 (Theorem of Smith) Let p be an odd prime and denote by C the set of points x = (x1, . . . , x,) satisfying 0 < xi < p (i = 1, ... , n). Let C* = C \ {(0, ... , 0)}. We define a box 13 in C as the set of all points x E C, which s a t i s f y 0 < vi < xi < vi + hi < p (i = 1, ... , n). Let f (X) be a polynomial in n variables (X = (X1, . , Xn), 0 < X i < p, i = 1, ... , n). Denote by N(13) the number of X E 13 for which f (X) = 0 (mod p). Then N(13) =
1 N(C) + 31
10
C,I
1] Sn(f, c)E.(13) cEC*
(4)
Algebraic Curves in Finite Geometry
211
where £c(13)
and SS(f, c) _ E e(2,ri/T4
ll
aEB
f (-)=O
where the sum extends over all x satisfying f (x) = 0, and c x denotes the ordinary inner product of c and x. Note that the Polya-Vinogradov inequality can be deduced from Smith's theorem with a slightly weaker remainder term: taking the box 8 = [1, h] x [1, p - 1] and the polynomial f (x, y) = x - y2 the number of quadratic residues in the interval [1, h] will just be half of N(8). The next lemmas clarify what the order of magnitude of N(8) is. The remainder term of (4) can be bounded using the following lemma.
Lemma 4.5 E(13) =
1]
IEc(13)I C Bpn . lognp,
(5)
cEC*
where B is an absolute constant depending only on n. For example if n = 2 and p > 60 then B = 1 can be supposed (see Section 3 of [42]). Lemma 4.6 (Lemma 1 of [91]) Let f (X) and zb(X, Y) be polynomials over GF(p), deg 0 = dl, deg f = d2, 1 < dl, d2 < p and suppose (X, Y) has no linear factors. Let e(2-ilP)'f(-)
S(f, b) _ *(-,Y)=0
where the summation is over all (x, y) E GF(p) x GF(p). Then S(f, 0) I < (di + 2d1d2 - 3d1)pl/2 + d2l.
(6)
So for the Polya-Vinogradov case we have d2 = 1, d1 = 2, hence (4) gives N(13) = h + log2 p(2 f + 4), so it is only a logp-factor weaker than the Polya-Vinogradov inequality. Actually, the extra logp-factor comes from Lemma 4.5. Using the fact that in our case one of the intervals of the box 13 is the complete residue system, the proof (see [42, Section 3, p. 293]) gives for n = 2 that JE(8)J < 2p2logp. This means that Smith's theorem almost gives back the Polya-Vinogradov inequality. The previous constructions produced relatively small complete arcs in the
sense that the largest arcs contain approximately q/2 points and the upper bound for the size of a complete arc is q + 1 or q + 2. Therefore it is surprising that in PC (2, q), q a fixed square, there is only one larger non-oval complete arc known. If q is not a square then the largest non-oval complete arc is the one containing half of the points of an elliptic cubic. If q is a square then
q2 +q+ 1 = (q+ J+ 1) (q - f + 1) and taking the orbit of a subgroup of order q - ,,Fq + 1 of a Singer group, one gets a (cyclic) arc of size q - vlq-+ 1. The arcs
212
T. Szonyi
are complete for q > 9. These arcs were first found by Kestenband [69] using a different construction, and later with this representation by Ebert [46], Fisher, Hirschfeld and Thas [50], and by Boros, Sz6nyi [30], see also Kestenband [70].
For exact values of m'(2, q) and n(2, q) in planes of small order, as well as information on the spectrum of complete arcs, see [59, 64, 99]. For minimal blocking sets the situation is similar to the one for complete
arcs. Bruen [33, 34] proved that the size of a non-trivial blocking set is at least q + Vq- + 1. On the other hand, Bruen and Thas [38] proved that a minimal blocking set contains at most q.,,,Fq + 1 points. Both bounds are sharp if q is a square; a subplane of order I-q (a Baer subplane) achieves the lower, a Hermitian arc (unital) achieves the upper bound. Recall that the Hermitian curve is the curve X3+1 + X2+1 + X3 +1 if q is a square. It has exactly qvlq- + 1 GF(q)-rational points and the lines of PG(2, q) meet this set in either 1 or Vq- + 1 points, see [58]. Hermitian arcs (or unitals) are combinatorial
analogues of Hermitian curves, that is sets of qlq- + 1 points in PG(2, q), q square, which intersect the lines in 1 or + 1 points. In particular, there is a unique tangent at each point. For further properties of Hermitian arcs and characterizations of Hermitian curves see [58]. However, it is not clear whether the upper bound is any good if q is not a square. In general, it would be desirable to know which are the possible sizes of minimal blocking sets, that is to determine the spectrum. This problem was posed for example in Cameron [41]. The next section contains some results in this direction for small blocking sets, that is for blocking sets whose size is at most 3(q + 1)/2. In particular, we shall see some improvements on Bruen's lower bound. Regarding large minimal blocking sets not much is known. For q =_ 1 (mod 4) there is a nice connection between maximal cliques in the Paley graph Pq and a particular class of minimal blocking sets obtained as the union of parabolas in a pencil. Denote by Pa, the parabola {(x, x2 + a) : x E GF(q)} U {(oo)}, where as usual (oo) is the infinite point of vertical lines.
Lemma 4.7 ([98]) Let A C GF(q), q =_ 1 (mod 4), and B = UaEA Pa. Then B is a minimal blocking set if and only if A is a maximal independent set in the Paley graph Pq. This immediately shows that for q - 1 (mod 4) there exist minimal blocking sets of size at least cq log q in PG(2, q), see [98], where the same result is obtained using Theorem 2.4 also for q - 3 (mod 4) with a more complicated construction of similar flavour. Actually, in both cases it is shown that there are minimal blocking sets of size cq log q. The construction idea of considering unions of conics was also used by Abbott and Liu [2] and Ughi [111]. The above lemma also means that the results on the cliques of the Paley graphs can be translated to results on minimal blocking sets. For example, if q is
an odd square then the independent set A = 8 GF(/), (6 is a non-square in GF(q)) gives a minimal blocking set of size q,,Fq + 1. This set is a nice example of a Hermitian arc which is not a Hermitian curve. Such Hermitian arcs
Algebraic Curves in Finite Geometry
213
were constructed by Buekenhout [40] and Metz [83] using 4-dimensional representations of projective planes of square order. For more on explicit planar representation of non-classical unitals, the reader is referred to the papers by Baker and Ebert [5, 6]. One can also repeat the trick used by Baker, Ebert, Hemmeter and Woldar
about halving a maximal clique and obtain various minimal blocking sets. This was done by Hirschfeld and Szonyi [65], who proved the following density theorem for planes of square order.
Theorem 4.8 For every A with 14 < A qo(A) there are minimal blocking sets B with c1g1+A < 1131
!5 CA
I+A.
(7)
The interval [q + 1, 3(q + 1)/2), that is small blocking sets, will be the subject of the next section. In the interval (3(q+ 1) /2, 2q) there are examples of minimal blocking sets of size 2q + 1 - d for every divisor d of q or q - 1. There are also other examples known, such as a minimal blocking set of size 19 in PG(2, 11); for more results, particularly for infinite series, see Gacs [51]. In the interval k E [2q, 3q - 3] we know almost everything: Innamorati and Maturo [67] and independently Illes, Szonyi and Wettl [66] constructed minimal blocking sets of size k for every such k. For a discussion of small planes, see [19] and [10].
Beyond this, almost nothing is known. Blokhuis and Metsch [22] proved that for q > 49, q square, there are no minimal blocking sets of size qvrq (that
is one less than the upper bound of Bruen and Thas). Some more (trivial) examples can be found in [66]. The situation is quite analogous for (k, n)-arcs. Barlotti [13] proved that for 1 < n < q + 1, k < qn - q + n and that equality can only occur when n I q. For q even Denniston [45] constructed (k, n)-arcs with k = nq - q + n in PG(2, q)
for every divisor n of q. A different construction was given by Thas [105]. Sometimes these arcs are called maximal arcs. Recently Ball, Blokhuis and Mazzocca [12] proved the long-standing conjecture that for q odd, equality in Barlotti's bound cannot occur; so there are no maximal arcs. A much shorter proof in the same style has been subsequently found by Ball and Blokhuis [11]. Previously this was only known for n = 3, see Cossu [44] and Thas [105]. Such
a hypothetical (2q + 3, 3)-arc would be a Steiner triple system embedded in PG(2, q). It is interesting to note that Thas' proof is based on the observation that in this Steiner triple system 3 non-collinear points generate an affine plane of order 3 (hence such a system is a so-called Hall triple system). This is proved using Segre's lemma of tangents (see the next section for Segre's lemma). Then Thas actually shows that the Steiner system must be an affine space over GF(3) and it gives the contradiction that both q and 2q + 3 must be powers of 3. If n is not a divisor of q, then the above bound by Barlotti was improved several times. Namely, if n is not a divisor of q then Lunelli and
214
T. Szonyi
Sce [79] proved that k < (n - 1)q + n - 3. Later they improved the bound to k < (n - 1)q + 8n/13, if n does not divide q and q is large compared to n. They also conjectured k < (n - 1)q + 1 if q is not divisible by n, but this was disproved by Hill and Mason [57]. On the other hand, if there is a line disjoint from the (k, n)-arc, then Blokhuis [18] verified the Lunelli-Sce conjecture; that is he proved k < (n - 1)q + 1 for such (k, n)-arcs. For large n > 2q/3, Hill [56] further improved the above bounds. On the constructive side not much is known. For small planes, we refer to Ball [9]. Besides the counterexamples by Hill and Mason, large (k, n)-arcs with k = (q - ph) (ph -1), n = q - ph (and q = p') were constructed by Mason [80, 81]. Hirschfeld and Szonyi [65] constructed (k, n)-arcs with k > (1 - e)nq, if n > ql/2+a which shows that Barlotti's bound is essentially correct. We used a pencil of conics and Theorem 2.4. Another approach in [65] used the PblyaVinogradov inequality for planes of prime order, and assumed n > c,,Fq log q. Let p be sufficiently large and let h be larger than vFp log p regarding its order of magnitude. Consider the set
K = {(x, x2 + a) : x E GF(p), 0 < a < h}. Elementary computation shows that it is enough to bound the intersection of K and horizontal lines; in other words, for a horizontal line t, : y = c to find the number of quadratic residues (squares) in the set {c - a 0 < a < h}. The Polya-Vinogradov inequality then gives that i6 n KI < h + 2.,,Fplogp. This immediately gives the asymptotic sharpness of Barlotti's bound if n is substantially larger than ,,Fplogp.
Proposition 4.9 Let p be a prime and suppose that n(p)/(\logp) --* oo. Then there is a (k, n')-arc in PG(2,p) with n(p) < n' < n(p) + 2,fp-log p and
k - pn' Choosing a smaller n and adding some points one can actually achieve that n' = n. Using the probabilistic method instead of the algebraic constructions, one can extend the result to n > c log q. This can be done by following essentially the proof of Erdos, Silverman and Stein [48]. Almost nothing is known for (k, n)-arcs with constant n > 2. Taking [n/2] disjoint conics, one can easily see that k can be at least [n/2](q + 1), but the upper bound is around (n - 1)q. Even for n = 3 the best example known is to take a non-singular cubic curve giving a (k, 3)-arc with k < q + 2vfq- + 1. Aart Blokhuis has offered some money, unfortunately in Hungarian Forints, to decide whether there are (k, 3)-arcs with k > (1 + E)q and to disprove the existence of (k, 3)-arcs with k > (2 - e)q. In Section 6 (k, n)-arcs for a divisor n of q are considered, if k is sufficiently close to the upper bound by Barlotti, that is for k > qn - q + n - E and e is small.
Algebraic Curves in Finite Geometry 5
215
The generalized Menelaus' theorem and applications: large arcs
The method discussed in this section is due to B. Segre [88, 90], and is one of the most successful theories in finite geometry. It consists of three parts. The first part is a nice result in classical algebraic geometry, the generalization
of Menelaus' theorem. The second part is crucial: it is a trick, sometimes called Segre's "Lemma of tangents" to associate algebraic curves to an arc in PG (2, q). The third and final step is to use Weil's theorem (or other bounds for the number of GF(q)-rational points of algebraic curves) to obtain upper bounds for the size of a complete are which is not an oval or hyperoval. Since
this theory is discussed in detail already in the first edition of Hirschfeld's book [58], I mainly state the results, and only consider one particular case to illustrate how the method works. At the end of this section further applications of Segre's method are collected. Let us fix a coordinate system in PG(2, q) and let Al : (1, 0, 0), A2 : (0, 1, 0), A3 : (0, 0, 1) be the fundamental triangle. If P is a point on one of the sides, then P can be written as (0, c, 1), (1, 0, c) or (c, 1, 0). The element c (=,A 0) is called the coordinate of P. The classical result of Ceva states that three points (0, c, 1), (1, 0, d) and (e, 1, 0) are collinear if and only if cde = -1. This can be generalized to curves of degree n.
Theorem 5.1 (B. Segre, [88]) Suppose that we are given a family G = {(0, ci,1), (1, 0, dj), (ek,1, 0) : i, j, k = 1, ... , n}
of 3n points and G fl {A1, A2, A3} = 0. Then there is a curve of degree n intersecting the sides of the triangle A1A2A3 in precisely the points of G, if and only if r 1i ci flj dj ilk ek = (-1)n.
Note that G is a multiset, the same point can occur more than once. If a point has multiplicity r in G, then we require that the curve of degree n intersects the corresponding side of A1A2A3 with multiplicity r. Actually, the proof is not difficult: the necessity comes from Vieta's formulas (for the product
of the roots of an equation), and the sufficiency is proved by a dimension argument. In most of the applications, the dual version of this theorem is needed. For this, we have to define the coordinate of a line passing through a vertex of the fundamental triangle. Such a line has equation X2 = dX3, X3 = dX1 or Xl = dX2; the element d is the coordinate of the line. The dual of Ceva's theorem is this: three lines through the vertices of the fundamental triangle are concurrent if and only if the product of their coordinates is 1. This is called Menelaus' theorem in classical projective geometry. The dual of the above generalization is the following.
Theorem 5.2 Suppose that we are given n lines (possibly with multiplicity), different from the sides of the fundamental triangle, through each vertex of the
216
T. Szonyi
fundamental triangle. This multiset of 3n lines is contained in an algebraic envelope of class n if and only if the product of the 3n coordinates is 1. An algebraic envelope is just a polynomial, the name underlines that its zeroes are considered as coordinates of lines and not points. For traditional reasons we use class for the degree of the polynomial defining an envelope. One can generalize further the above theorems, instead of 3 lines or points one can consider k lines or points. This generalization will be stated only in the dual form.
Theorem 5.3 (Segre's generalized Menelaus' theorem) Suppose that we have
a k-arc and at each point we are given a multiset of n lines, which are all tangents to the k-arc. Then these kn lines are contained in an algebraic envelope of class n if and only if for every 3 points of our k-arc the corresponding 3n lines are contained in an algebraic envelope of class n. So if there is an envelope locally for every triangle containing the 3n lines, then there is an envelope globally containing the kn lines. The condition that the k points form a k-arc can be relaxed. Since only the original version of Segre's generalized Menelaus' theorem will be used later, we do not state this even more general version, just refer to the paper by Blokhuis, Cameron and Thas [21].
Note that the famous theorem of Segre, saying that the (q + 1)-arcs in PG(2, q) (q odd) are conics, was proved using this method implicitly. Then Segre's idea was to embed arcs with size close to q in ovals or hyperovals. At the end of this section we shall make it clear what "close" means here. For q even it is a nice combinatorial exercise to prove that (q + 1)-arcs are never complete;
they are always contained in a (q + 2)-arc (that is a hyperoval). Indeed, there is a unique tangent at each point, and the q+1 tangents cover all points; hence these line have to form a pencil. However, for the embedding of q-arcs there is no combinatorial proof; in fact, Menichetti [82] has constructed complete q-arcs in some non-desarguesian planes. Let us illustrate Segre's method by proving the incompleteness of q-arcs in desarguesian planes of even order, a result first proved by Tallini [103].
Theorem 5.4 (Tallini) In PG(2, q), q even, q-arcs are never complete. Proof Let Q be a q-arc. Through each point of Q there are exactly 2 tangents to the q-arc Q at that point, so in total there are 2q tangents. We wish to use Segre's generalized Menelaus' theorem for this set of 2q lines. According to the theorem, we only have to check the condition for triangles contained in Q. Let Al, A2, A3 be three distinct points of Q and choose coordinates in such a way that Al = (1, 0, 0), A2 = (0, 1, 0), A3 = (0, 0, 1). Form a matrix of size (q - 3) x 3 whose rows are indexed by the points P E Q \ {Al, A2, A3} and the columns by Al, A2 and A3. Put the coordinate of the line AiP in the position
Algebraic Curves in Finite Geometry
217
(P, Ai). Then Menelaus' theorem implies that the product of the elements in one row is just 1. Thus the product of the elements in this matrix is 1q-3, computing it row by row. Extend the matrix to a (q - 1) x 3 matrix by adding two rows. In these rows we write the coordinates of the two tangents at the points Al, A2 and A3. In any column of this extended matrix each element of GF(q) occurs exactly once, hence their product is 1. Therefore, the product of all elements in the extended matrix is 13, computing it column by column. Hence the product of the coordinates of the tangents at Al, A2 and A3 is 1, and Menelaus' theorem (for a triangle) implies that they are contained in an envelope of class 2. From the general theorem of Menelaus it also follows that this envelope of class two does not depend on the points Al, A2, A3. The total number of tangents is 2q, hence our curve contains at least 2q GF(q)rational points. An irreducible conic contains q + 1 points, hence our curve is the union of two pencils, if q > 4. Geometrically this means that there are two points N, N' with the property that joining N or N' to any point of Q yields a tangent to Q. In other words, either of the two points can be added to Q. Hence Q cannot be complete. Note that this also implies, by the combinatorial argument showing the incompleteness of (q + 1)-arcs, that Q can be embedded
in a hyperoval by adding N and N' one after the other. In particular, this means that the line joining N and N' does not intersect Q. In general, one can apply Segre's Lemma of tangents to the set of all tangents to a k-arc and show that the tangents are contained in an envelope of
class q + 2 - k or 2(q + 2 - k) according as q is even or odd. This is the cornerstone of the proof of almost all embedding theorems for large arcs. For a detailed list of properties of these algebraic envelopes, see Hirschfeld [58, Theorems 10.3.1 and 10.4.1]. In the general case the envelope rt or Flt contains at least kt = k(q+2-k) GF(q)-rational points and a similar (but slightly more complicated) lower bound holds for the number of points of any component of it. Comparing this with the Weil bound gives a lower bound for the degree of non-linear components. The other possibility is to use the Weil bound for every component and adding these bounds up gives an upper bound for the number of GF(q)rational points of rt or Flt. The linear components, as in the above example, correspond to points that can be added to the arc, hence the method yields an upper bound for the size of a complete arc which is not an oval or hyperoval. This strategy is due to B. Segre [88, 90] and gives an upper bound for m'(2, q). In the final step of Segre's method one can use refinements and variants of Weil's theorem. Since most of them are technically difficult, we only mention an elementary upper bound for arbitrary q, and the Stohr-Voloch bound for q prime. The first observation was used by Thas in [106].
Let C be a curve of degree d defined over GF(q), which does not have a
linear component defined over GF(q). Then it has at most N < qd - q + d points over GF(q).
T. SzSnyi
218
The proof is easy: such a curve is a (N, d')-arc for some d' < d and Barlotti's bound (see Section 4) immediately gives the upper bound. The Stohr-Voloch theorem is a bound which is in most cases stronger than Weil's bound. For an exposition, see Hirschfeld [61]. We state it in the simplest case, when there are no technical conditions. Let C be an absolutely irreducible plane curve defined over GF(p), p prime.
If 3 < n < p/2 and C has s double points, then
N < 2n(5(n - 2) + p - 10s),
(8)
where again N denotes the number of GF(p)-rational points of C.
The best results, obtained with Segre's strategy, and using the various upper bounds for the number of rational points, are the following.
Theorem 5.5 For the size of the second largest complete arc, we have:
(a) m'(2, q) < q - Fq + 1, if q is an even square (Segre [90], Thas 106]), (b) m'(2, q) < q - 4 + 16, if q is an odd square (Thas [107]),
(c) m(2, q) < q-
pq4
+
1629p
+1,ifq=p"`,pprime, p>2,n>3,nodd
(Voloch [114]),
(d) m'(2, q) < q (e) M'(2, q)
2 (Voloch (113]), 45
(f) M'(2, q) < q -
2
+ 5, if q = ph, p > 5, (Hirschfeld, Korchmaros [62]),
(g) m'(2, q) < q - Lq + 3, for q = ph with p > 5, if q > 192 and q
52,
(Hirschfeld, Korchmaros [63]).
Let us see that the Stohr-Voloch bound in the prime case indeed gives the
bound in (e). The envelope containing the tangents has class 2(p + 2 - k). Take an irreducible component of class n. As in [58], n > 3 can be assumed. The Stohr-Voloch bound (see (8)) gives the inequality
2(p+2-k) < 5n(5(n-2)+p),
Algebraic Curves in Finite Geometry
219
which implies that 2n - 4 > p/10 + 1 - k/2. Using n < 2k it gives k < (44p + 40) /45 indeed. Similarly, the above mentioned elementary upper bound of Thas gives the
bound in case (a). Note that the arcs constructed by Kestenband [69], Ebert [46], Fisher, Hirschfeld and Thas [50] and Boros, Szonyi [30] (see Section 4), show that for q square q + 1 < m'(2, q). In particular, part (a) is sharp. On the other hand, if q is not a square then the best lower bound known for m'(2, q) is only [q + 2,]/2, coming from Zirilli's construction (see Theorem 4.3). Let us close this section with further applications of Segre's curve-making method. The first one is about sets having a large number of internal nuclei. Let K be a k-set in PG(2, q). A point P E K is called an internal nucleus of K if each line through P meets K in at most two points including P. This notion was introduced for k = q + 2 by Bichara and Korchmaros [14] and generalized by Wettl [120]. At an internal nucleus P a line t is a tangent to K if it meets K in just P. From our point of view the main result of Wettl is the following: for q odd the tangents to K at the internal nuclei are contained in an algebraic envelope of class 2(q + 2 - k). Of course, the proof is based on Segre's Lemma of tangents. Wettl [120] used the theorem for characterizing (q + 1)-sets with at least 5 nuclei and collinear non-nuclei. Therefore, the situation is quite analogous to the case of arcs, the only difference is that the curve does not contain too many points, since typically at most (q + 1)/2 points can be nuclei. This is the content of the next result, see Szonyi [97].
Theorem 5.6 Let K be a k-set with k > q - //8 + c (q odd, q > qo(c)). If K is not an arc, then it contains at most (q + 1)/2 internal nuclei. The next application is due to Blokhuis, Seress and Wilbrink: a set S is a set without tangents if every line either is disjoint from S or meets S in at least 2 points. The main result of [25] is the following.
Theorem 5.7 (Blokhuis, Seress, Wilbrink) Let S be a set without tangents in PG(2, q), q odd. Then ISI > q + 2q/4 + 2. The proof is a clever use of Segre's lemma of tangents. Using the generalized Menelaus' theorem (Theorem 5.3) and Bezout's theorem to glue together smaller curves, one obtains an envelope of class 2(ISI - q - 2), containing the lines that intersect S in more than two points. The inequality is obtained by using the upper bound for the number of singular points of a curve. Let me finally mention the result, which led to the Blokhuis, Cameron, Thas [21] generalization of Segre's generalized Menelaus' theorem. This is due to Thas [108], and gives the following elegant characterization of Hermitian curves:
A Hermitian arc H is a Hermitian curve if and only if tangents of H at collinear points of H are concurrent.
220
6
T. Szonyi
The Redei polynomial and applications: blocking sets
and (k, n)-arcs In this section a new approach of associating curves to blocking sets and (k, n)-arcs will be discussed. We shall follow the papers [23, 100, 101]. The first paper in which algebraic curves were used to prove results on blocking sets was the paper by Blokhuis, Pellikaan and Szonyi [23]. Then in [100] a pair of curves was introduced and the bounds for the size of a minimal blocking set were obtained by using Bezout's theorem applied for this pair of curves. Before sketching this new approach, let us summarize what was known for blocking sets in PG(2, q). For a more detailed account, together with historical remarks, see Blokhuis [19]. The first non-trivial result for blocking sets is due to Bruen and Pelikan;t Bruen [33, 34] proved that the size of a non-trivial blocking set is at least q + Vrq- + 1. This is sharp for q square (see [33] and also Section 4). For q a non-square the bound was improved several times, the most recent one is Blokhuis's theorem. Theorem 6.1 (Blokhuis [17, 19]) If q is a prime, then the size of a non-trivial blocking set is at least 3(q + 1)/2. If q = ph is neither a square nor a prime, then the size of a non-trivial blocking set is at least q + pq + 1.
Note that the bound in the prime case is sharp (see [77]) and it solved a thirty-year old conjecture of Jane di Paola [84]. The bound is also sharp in the case q = p3; it solved a 25-year old conjecture of Bruen [34]. We shall return to the examples later. Let B be a blocking set of PG(2, q). A point P E B is called essential if B \ {P} is not a blocking set. The blocking set B is minimal (or irreducible) if and only if every P E B is essential. Geometrically this means that through each point of the blocking set B there is a line intersecting B in just one point. According to the standard terminology introduced in Section 4 such a line will be called a tangent and, more generally, a line intersecting B in r points will be called an r-secant (or a line of length r). Let L be the line at infinity, and suppose that {(oo)} = L fl B, that is L is a tangent to B. Give affine coordinates to the points of U := B \ L; namely,
letU={(ai,bi):i=1,...,q+k}. So 1131 =q+k+l. Definition The Redei-polynomial of U is defined as follows:
H(X, Y) := fJ(X + aiY - bi) = Xq+k + h1(Y)Xq+k-1 + ... + hq+k(1') (9) i
Note that deg(hj) < j for j = 1, ... , q+k. If H(X, Y) is considered for a fixed Y = y as a polynomial of X, then we write Hy(X) (or just H(X, y)). tPelikan and Pellikaan are different people
Algebraic Curves in Finite Geometry
221
Definition Let C be the afline curve of degree k defined by k
f (X, Y) = X + hi (Y)X
k-1
+ ... + hk (y).
(10)
By the remark for the degree of hj in the previous definition, f (X, Y) has degree k. The next proposition summarizes some important properties of the Redei polynomial and of these curves.
Theorem 6.2 (1) For a fixed (y) E L \ B the polynomial (X9 - X) divides Hy(X). Moreover, if k < q - 1 then Hy(X)/(X9 - X) = f (X, y) for every (y) E L \ B; and f (X, y) splits into linear factors over GF(q) for these fixed y's.
(2) For a fixed (y) E L \ B, the element x is an r-fold root of Hy (X) if and only if the line with equation Y = yX + x intersects U in exactly r points. If the line with equation Y = y meets f (X, Y) at (x, y) with multiplicity m, then the line with equation Y = yX + x meets U in exactly m + 1 points.
The first part of this theorem shows that f has a lot of GF(q)-rational points, the second part helps us translate geometric properties of U into properties of f. The next lemma shows that the linear components of f correspond to points of B which are not essential.
Lemma 6.3 If a point P = (a, b) E B is not essential, then X +aY - b divides f (X, Y) (as polynomials in two variables). Conversely, if k < q and X + aY - b divides f (X, Y), then (a, b) E B and (a, b) is not essential. It will be convenient to suppose that not only the line at infinity but also
the y-axis is a tangent to the blocking set B. Since the Redei polynomial H(X, Y) vanishes for all (x, y) E GF(q) x GF(q) (see Theorem 6.2(1)), we can write it as
H(X, Y) = (X9 - X) f (X, y) + (Y9 - Y)g(X, y), where deg(f), deg(g) < k as polynomials in two variables (see [2, 9]). Note that f here is the same as the one in (10). If one fixes Y = y then H(X, y) is divisible by (X9 - X) and for an (x, y) E GF(q) x GF(q) we have that f (x, y) = 0 if and only if the line with equation Y = yX + x intersects U in at least two points (cf. Theorem 6.2(2)). The next theorem summarizes some important properties of the pair of curves (f, g).
Theorem 6.4 The curves f (x, y) and g(x, y) have the same GF(q)-rational affine points, but they do not have common components.
The next result, already used in [23], gives a lower bound for the number of GF(q)-rational points on a component of f.
222
T. Szonyi
Theorem 6.5 (Blokhuis, Pellikaan, Szonyi) Let h be a component off with h' 0 0. Let s denote the degree of h. Then the number of GF(q)-rational points of h is at least qs - s(s - 1). Using these properties of the curves f and g and the above lower bound for the number of rational points, one can show that for small blocking sets,
that is when k < (q + 1)/2, every component of f must have zero partial derivative with respect to X. Geometrically, this means the following: If B is a blocking set of size less than 3(q + 1)/2, then each line intersects it in 1 modulo p points. In particular, this gives a new proof of Blokhuis' famous lower bound IBI > 3(p+ 1)/2 (see Theorem 6.1) for blocking sets in PG(2,p), p prime.
To say more in the case q = p", n > 1, components with zero partial derivative should be studied in detail and the following theorem can be proved.
Theorem 6.6 (Szonyi [100]) Let B be a minimal blocking set in PG(2, q), q = p". Suppose that IBI < 3(q + 1)/2. Then
5 it is not possible that n/2 > e > n/3. On the other hand, using the fact that each line intersects our blocking sets in 1 modulo pe points, also the upper bound can slightly be improved. These
T. Szonyi
224
observations are essentially due to Blokhuis and Polverino [24], who obtained the following theorem.
-jj
Theorem 6.8 (Blokhuis, Polverino) With the notation of Theorem 6.6, < q+1+pe q/pe+ll pe+l
1131
< -1+
(pe + 1) (q + 1) 2
where A = (1 + (pe + 1)(q + 1))2 - 4(pe + 1)(q2 + q + 1).
Note that this new upper bound is asymptotically q+q/pe+q/pee+2q/p3e+ This improvement can be used to determine the possible sizes of minimal blocking sets in PG(2,p3) as the following corollary shows.
Corollary 6.9 (Blokhuis, Polverino) Let B be a non-trivial minimal blocking set in PG(2,p3). Then CBI =p3+p2+1 orp3+p2+p+1 or 1131 > 3(p3+1)/2. Again, the result extends immediately to blocking sets with e = n/3. Note that there are examples of minimal blocking sets (of Redei type) for both cardinalities. The same approach using algebraic curves can also be used to prove the affine blocking set theorem of Jamison [68] and Brouwer-Schrijver [31], and its generalization for multiple blocking sets, due to Bruen [36]. Concerning these problems, the reader is referred to Blokhuis' papers [18] and [19]. Using the previous results on blocking sets one can improve on the LunelliSce bound for the cardinality of a complete arc, since the secants of a complete arc form a blocking set in the dual plane.
Theorem 6.10 (Blokhuis, Ball, Blokhuis-Polverino) Let K be a complete karc in PG(2, q), and assume that q = p, p2 or p3, where p is a prime. Then k > 3q.
Regarding multiple blocking sets many fewer results and examples are known. Taking the union of s disjoint Baer subplanes gives an s-fold blocking set, which will be characterized by the next theorem. Using the lacunary polynomial approach, Ball [9] (for the prime case), and Blokhuis, Storme and Szonyi [26] proved the following lower bounds. Note that IBS > s(q + 1) is obvious.
Theorem 6.11 ([9, 26]) Let B be an s-fold blocking set in PG(2, q) of size
s(q+1)+c. Let c2=c3=2-1/3 andcp=1 forp>3. 1. (Ball) Ifq = p prime ands < (q-1)/2, then c > (q+l)/2. Ifs > (q+l)/2 then s + c > q. 2. Ifq = p2d+1 and s < q/2 - cpg213, then c >
cpg213.
Algebraic Curves in Finite Geometry
225
3. If 4 < q is a square, s < q'/4/2 and c < cpg2I3, then c > s.,Fq and B contains the union of s disjoint Baer subplanes.
4. If q = p2, s < g1/4/2 and c < p 14 +
(p-+1) /2], then c > sV14 and B contains the union of s disjoint Baer subplanes.
The special cases s = 2, 3, that is the cases of double and triple blocking sets, were proved earlier by Ball and Blokhuis [10] and Ball [8]. Note that the case s = 1 yields considerable improvement on the bounds of Blokhuis mentioned in Theorem 6.1. It is worthwhile to mention that for s = (q - 1)/2 the set of external points of a conic shows that Ball's bound is sharp. For s = (q + 1)/2 the same set together with all points of the conic but one is an example. For small s the bounds do not seem to be sharp. In fact, Blokhuis
conjectures that for s = 2 and q = p prime c > q - 2. Still for s = 2 the bound does not seem to be exact, if q is not a square. Recently, L. Lovasz and the author tried to extend the above method, that is the use of curves, to multiple blocking sets. To an s-fold blocking set is associated a set of s + 1 curves with almost the same set of rational points. Since this research is still in progress let me only mention an improvement on the above bounds in the special case s = 2, q = p3. This indicates that the use of curves can give more than the lacunary polynomial approach also in case of multiple blocking sets. For a general s, it is not yet clear whether the method using curves gives the extra term s in the bound or not.
Proposition 6.12 (Lovasz, Szonyi [78]) A double blocking set of PG(2, p3) has size at least 2(p3 + p2 + 1 - 3p).
Now let us turn to (k, n)-arcs. Regarding the bounds for k we refer to Section 4. In the manuscript [101] the present author shows that for n = p, q = ph, a (k, n)-arc K with k > qn - q + n - e points, e < c can be embedded in a maximal arc. This result can probably be extended to any divisor n of q, if ne + e'4 is smaller than c q. The embeddability was only known before for e = 1 (Thas [105]) and even for s = 2 only partial results were known (Wilson [121]).
Theorem 6.13 ([101]) In PG(2,q), q = ph let K be a (k,p)-arc with k > qp-q+p-cg114. Then K can be embedded in a (k, p)-arc with k = qp-q+p. Since maximal arcs do not exist for p > 2 (see [12]), the above result simply means that for a (k, p)-arc in PG (2, q), q = ph, p > 2, we have k
6) even, Ars Combinatoria, 16 (1983), 103-111.
[2] H. L. Abbott & A. Liu, Property of B(s) and projective planes, Ars Combinatoria, 20 (1985), 217-220. [3] N. Alon, Tools from higher algebra, Chapter 32, in Handbook of Combinatorics (eds. R. L. Graham, M. Gr6tschel & L. Lovasz), North-Holland, Amsterdam (1995), pp. 1749-1783.
[4] L. Babai, A. Gal & A. Wigderson, Superpolynomial lower bounds for monotone span programs, submitted to Combinatorica. [5] R. D. Baker & G. L. Ebert, On Buekenhout-Metz unitals of even order, European Journal of Combinatorics, 13 (1992), 109-117.
[6] R. D. Baker & G. L. Ebert, On Buekenhout-Metz unitals of odd order, Journal of Combinatorial Theory, Series A, 60 (1992), 67-84.
[7] R. D. Baker, G. L. Ebert, J. Hemmeter & A. Woldar, Maximal cliques in the Paley graph of square order, Journal of Statistical Planning and Inference, 56 (1996), 33-38.
[8] S. M. Ball, On the size of a triple blocking set in PG(2, q), European Journal of Combinatorics, 17 (1996), 427-435. [9]
S. M. Ball, Multiple blocking sets and arcs in finite planes, Journal of the London Mathematical Society (2), 54 (1996), 581-593.
[10] S. M. Ball & A. Blokhuis, On the size of a double blocking set in PG(2, q), Finite Fields and their Applications, 2 (1996), 125-137.
[11] S. M. Ball & A. Blokhuis, An easier proof of the maximal arcs conjecture, manuscript, 1996.
[12] S. M. Ball, A. Blokhuis & F. Mazzocca, Maximal arcs in PG(2, q), q odd do not exist, Combinatorica, in press.
Algebraic Curves in Finite Geometry
229
[13] A. Barlotti, Su {k; n}-archi di un piano lineare finito, Bollettino della Unione Matematica Italiana, 11 (1956), 553-556. [14] A. Bichara & G. Korchmaros, n2-sets in a projective plane which determine exactly n2 + n lines, Journal of Geometry, 15 (1980), 175-181. [15] A. Blokhuis, On subsets of GF(q 2) with square differences, Indagationes Mathematicae, 46 (1984), 369-372.
[16] A. Blokhuis, Polynomials in finite geometry and combinatorics, in Surveys in Combinatorics, 1993 (ed. K. Walker), London Mathematical Society Lecture Note Series, 187, Cambridge University Press, Cambridge (1993), pp. 35-52.
[17] A. Blokhuis, On the size of a blocking set in PG(2,p), Combinatorica, 14 (1994), 273-276.
[18] A. Blokhuis, On multiple nuclei and a conjecture of Lunelli and Sce, Bulletin of the Belgian Mathematical Society, Simon Stevin, 3 (1994), 349-353.
[19] A. Blokhuis, Blocking sets in Desarguesian planes, in Combinatorics, Paul Erdds is Eighty, Volume 2 (eds. D. Miklos, V.T. Sos and T. Szonyi), Bolyai Society Mathematical Studies, 2, Bolyai Society, Budapest (1996), pp. 133-155.
[20] A. Blokhuis, S. M. Ball, A. E. Brouwer, L. Storme & T. Szonyi, On the number of slopes determined by a function on a finite field, manuscript, 1996.
[21] A. Blokhuis, P. J. Cameron & J. A. Thas, On a generalization of a theorem of B. Segre, Geometriae Dedicata, 43 (1992), 299-305. [22] A. Blokhuis & K. Metsch, Large minimal blocking sets, strong representative systems and partial unitals, in Finite Geometry and Combinatorics
(eds. F. De Clerck et al.), London Mathematical Society Lecture Note Series, 191, Cambridge University Press, Cambridge (1993), pp. 37-52. [23] A. Blokhuis, R. Pellikaan & T. Szonyi, Blocking sets of almost Redei type, Journal of Combinatorial Theory, Series A, in press. [24] A. Blokhuis & O. Polverino, private communication, 1996.
[25] A. Blokhuis, A. Seress & H. A. Wilbrink, On sets of points without tangents, Mitteilungen aus dem Mathematischen Seminar Giessen, 201 (1991), 39-44.
[26] A. Blokhuis, L. Storme & T. Szonyi, Multiple blocking sets and Baersubplanes, manuscript, 1995.
T. Szonyi
230
[27] B. Bollobas, Geodesics in oriented graphs, Annals of Discrete Mathematics, 20 (1984), 67-73. [28] B. Bollobas, Random Graphs, Academic Press, New York (1985).
[29] B. Bollobas & A. Thomason, Graphs which contain all small graphs, European Journal of Combinatorics, 2 (1981), 13-15. [30] E. Boros & T. Szonyi, On the sharpness of a theorem of B. Segre, Combinatorica, 6 (1986), 261-268.
[31] A. E. Brouwer & A. Schrijver, The blocking number of an affine space, Journal of Combinatorial Theory, Series A, 24 (1978), 251-253.
[32] A. E. Brouwer & H. A. Wilbrink, Blocking sets in translation planes, Journal of Geometry, 19 (1982), 200. [33] A. A. Bruen, Baer subplanes and blocking sets, Bulletin of the American Mathematical Society, 76 (1970), 342-344.
[34] A. A. Bruen, Blocking sets in finite projective planes, SIAM Journal on Applied Mathematics, 21 (1971), 380-392.
[35] A. A. Bruen, Arcs and multiple blocking sets, in Combinatorica, Symposia Mathematica, 28, American Mathematical Society, Providence, Rhode Island (1986), pp. 15-29. [36] A. A. Bruen, Polynomial multiplicities over finite fields and intersection sets, Journal of Combinatorial Theory, Series A, 60 (1992), 19-33. [37] A. A. Bruen & R. Silverman, Arcs and blocking Sets II, European Journal of Combinatorics, 8 (1987), 351-356.
[38] A. A. Bruen & J. A. Thas, Blocking sets, Geometriae Dedicata, 6 (1977), 193-203.
[39] A. A. Bruen, J. A. Thas & A. Blokhuis, On M.D.S. codes, arcs in PG(n, q) with q even, and a solution of three fundamental problems of B. Segre, Inventiones Mathematicae, 92 (1988), 441-459. [40] F. Buekenhout, Existence of unitals in finite translation planes of order q2 with kernel of order q, Geometriae Dedicata, 5 (1976), 189-194.
[41] P. J. Cameron, Four lectures on projective geometry, in Finite Geometry (eds. C. A. Baker & L. M. Batten), Lecture Notes in Pure and Applied Mathematics, 103, Marcel Dekker, New York (1985), pp. 27-65.
[42] J. H. H. Chalk, The number of solutions of congruences in incomplete residue systems, Canadian Journal of Mathematics, 15 (1963), 291-296.
Algebraic Curves in Finite Geometry
231
[43] F. R. K. Chung, R. L. Graham & R. M. Wilson, Quasi-random graphs, Combinatorica, 9 (1989), 345-362.
[44] A. Cossu, Su alcune propriety dei {k, n}-archi di un piano proiettivo sopra un corpo finito, Rendiconti di Matematica e delle sue Applicazioni, 20 (1961), 271-277.
[45] R. H. F. Denniston, Some maximal arcs in finite projective planes, Journal of Combinatorial Theory, 6 (1969), 217-219.
[46] G. L. Ebert, Partitioning projective geometries into caps, Canadian Journal of Mathematics, 37 (1985), 1163-1175.
[47] P. Erdos, On a problem in graph theory, Mathematical Gazette, 47 (1963), 220-223.
[48] P. Erdos, R. Silverman & A. Stein, Intersection properties of families of sets of nearly the same size, Ars Combinatoria, 15 (1983), 247-259.
Erdos & J. Spencer, Probabilistic Methods in Combinatorics, Akademiai Kiadd and Academic Press, Budapest and New York (1974).
[49] P.
[50] J. C. Fisher, J. W. P. Hirschfeld & J. A. Thas, Complete arcs in planes of square order, Annals of Discrete Mathematics, 30 (1986), 243-250.
[51] A. Gacs, On the number of directions determined by a pointset in AG(2, p), submitted to Discrete Mathematics.
[52] R. L. Graham & J. Spencer, A constructive solution to a tournament problem, Canadian Mathematical Bulletin, 14 (1971), 45-48. [53] J. van der Geer & J. H. van Lint, Introduction to Coding Theory and Algebraic Geometry, Birkhauser, Basel (1988).
[54] V. D. Goppa, Geometry and Codes, Kluwer, Dordrecht (1988).
[55] E. Hadnagy, Small complete arcs in Galois-planes of prime order, manuscript, 1996. [56] R. Hill, Some problems concerning (k, n)-arcs in finite projective planes, Rendiconti del Seminario Matematico di Brescia, 7 (1984), 367-383.
[57] R. Hill & J. R. M. Mason, On (k, n)-arcs and the falsity of the LunelliSce conjecture, in Finite Geometries and Designs (eds. P. J. Cameron, J. W. P. Hirschfeld & D. R. Hughes), London Mathematical Society Lecture Note Series, 49, Cambridge University Press, Cambridge (1980), pp. 153-168.
232
T. Szonyi
[58] J. W. P. Hirschfeld, Projective Geometries over Finite Fields, Clarendon Press, Oxford (1979). 2nd edition, in press.
[59] J. W. P. Hirschfeld, Maximum sets in finite projective spaces, in Surveys in Combinatorics (ed. E. K. Lloyd), London Mathematical Society Lecture Note Series, 82, Cambridge University Press, Cambridge (1983), pp. 55-76. [60] J. W. P. Hirschfeld, The Weil conjectures in finite geometry, in Combinatorial Mathematics X, Lecture Notes in Mathematics, 1036, Springer, Berlin (1983), pp. 6-23. [61] J. W. P. Hirschfeld, Algebraic Curves, Arcs, and Caps over Finite Fields, Quaderni del Dipartimento di Matematica dell'Universita di Lecce, Q.5, (1986).
[62] J. W. P. Hirschfeld & G. Korchmaros, On the embedding of an arc into a conic in a finite plane, Finite Fields and their Applications, 2 (1996), 274-292.
[63] J. W. P. Hirschfeld & G. Korchmaros, The number of rational points on an algebraic curve over a finite field, manuscript, 1996.
[64] J. W. P. Hirschfeld & L. Storme, The packing problem in statistics, coding theory and finite geometry, Journal of Statistical Planning and Inference, in press. [65] J. W. P. Hirschfeld & T. Szonyi, Constructions of large arcs and blocking sets in finite planes, European Journal of Combinatorics, 12 (1991), 499511.
[66] T. Illes, T. Szonyi & F. Wettl, Blocking sets and maximal strong representative sytems in finite projective planes, Mitteilungen aus dem Mathematischen Seminar Giessen, 201 (1991), 97-107.
[67] S. Innamorati & A. Maturo, On irreducible blocking sets in projective planes, Ratio Mathematica, 2 (1991), 151-155. [68] R. Jamison, Covering finite fields with cosets of subspaces, Journal of Combinatorial Theory, Series A, 22 (1977), 253-266. [69] B. C. Kestenband, Unital intersections in finite projective planes, Geometriae Dedicata, 11 (1981), 107-117. [70] B. C. Kestenband, A family of complete arcs in finite projective planes, Colloquium Mathematicum, LVII (1989), 59-67. [71] J. H. Kim & V. H. Vu, private communication.
Algebraic Curves in Finite Geometry
233
[72] J. Kollir, L. Ronyai & T. Szabo, Norm-graphs and bipartite Turin numbers, Combinatorica, 16 (1996), 399-406.
[73] G. Korchmiros, New examples of k-arcs in PG(2, q), European Journal of Combinatorics, 4 (1983), 329-334.
[74] T. Koviri, V. T. Sos & P. Turin, On a problem of K. Zarankiewicz, Colloquium Mathematicum, 3 (1954), 50-57.
[75] L. Lombardo-Radice, Sul problema dei k-archi completi di 52,9, Bollettino dell'Unione Matematica Italiana, 11 (1956), 178-181.
[76] R. Lidl & H. Niederreiter, Finite Fields, Encyclopedia of Mathematics and its Applications, 20, Addison-Wesley, Waltham (1983). [77] L. Lovisz & A. Schrijver, Remarks on a theorem of Redei, Studia Scientiarum Mathematicarum Hungarica, 16 (1981), 449-454.
[78] L. Lovisz & T. Sz6nyi, Multiple blocking sets and algebraic curves, manuscript, 1996.
[79] L. Lunelli & M. Sce, Considerazioni aritmetiche e risultati sperimentali sui {K, n}q archi, Rendiconti Istituto Lombardo, Accademia di Scienze e Lettere, A, 98 (1964), 3-52.
[80] J. R. M. Mason, On the maximum sizes of certain (k, n)-arcs in finite projective geometries, Mathematical Proceedings of the Cambridge Philosophical Society, 91 (1982), 153-169.
[81] J. R. M. Mason, A class of ((pn - pm)(pn - 1) pn _ pm)-arcs in PG(2, q), Geometriae Dedicata, 15 (1984), 355-361.
[82] G. Menichetti, q-archi completi nei piani di Hall di ordine q = 2', Atti dell'Accademia Nazionale dei Lincei, Classe di Scienze Fisiche Matematiche e Naturali, Rendiconti, 56 (1974), 518-525. [83] R. Metz, On a class of unitals, Geometriae Dedicata, 8 (1979), 125-126.
[84] J. di Paola, On minimum blocking coalitions in small projective plane games, SIAM Journal on Applied Mathematics, 17 (1969), 378-392. [85] L. Redei, Liickenhafte Polynome Fiber endlichen Korpern, Akademiai Kiado and Birkhauser Verlag, Budapest and Basel (1970). [86] R. Schoof, Non-singular plane cubic curves over finite fields, Journal of Combinatorial Theory, Series A, 46 (1987), 183-211.
[87] B. Segre, Ovals in a finite projective plane, Canadian Journal of Mathematics, 7 (1955), 414-416.
234
T. Szonyi
[88] B. Segre, Le geometrie di Galois, Annali di Matematica Pura ed Applicata, 48 (1959), 1-97. [89] B. Segre, Ovali e curve a nei piani di Galois di caratteristica due, Atti dell'Accademia Nazionale dei Lincei, Classe di Scienze Fisiche Matematiche e Naturali, Rendiconti, 32 (1962), 785-790. [90] B. Segre, Introduction to Galois geometries (edited by J. W. P. Hirschfeld), Atti dell'Accademia Nazionale dei Lincei, Classe di Scienze Fisiche Matematiche e Naturali, Memorie, 8 (1967), 133-236. [91] R. A. Smith, The distribution of rational points on hypersurfaces defined over a finite field, Mathematika, 17 (1970), 328-332.
[92] K. O. Stohr & J. F. Voloch, Weierstrass points and curves over finite fields, Proceedings of the London Mathematical Society, 52 (1986), 1-19.
[93] L. Storme, k-arcs in PG(n, q) and linear M.D.S. codes, Academiae Analecta, Belgium, 55 (1993), 88-126.
[94] T. Szonyi, Small complete arcs in Galois planes, Geometriae Dedicata, 18 (1985), 161-172.
[95] T. Szonyi, Note on the order of magnitude of k for complete k-arcs in PG(2, q), Discrete Mathematics, 66 (1987), 263-266.
[96] T. Szonyi, Arcs in cubic curves and 3-independent subsets of abelian groups, in Combinatorics, Eger, Colloquia Mathematica Societatis Janos Bolyai, 52, North-Holland, Amsterdam (1987), pp. 499-508.
[97] T. Szonyi, k-sets in PG(2, q) having a large set of internal nuclei, in Combinatorics '88, Volume 2 (eds. A. Barlotti et al.), Mediterranean Press, Rende (1991), pp. 449-458. [98] T. Sz6nyi, Note on the existence of large minimal blocking sets in Galois planes, Combinatorica, 12 (1992), 227-235.
[99] T. Szonyi, Arcs, caps, codes and 3-independent subsets, in Giornate di Geometrie Combinatorie (eds. G. Faina & G. Tallini), University di Perugia, Perugia (1993), pp. 57-80.
[100] T. Sz6nyi, Blocking sets in desarguesian affine and projective planes, Finite Fields and their Applications, in press. [101] T. Szonyi, On the embeddability of (k, p)-arcs, manuscript, 1996. [102] E. Szekeres & G. Szekeres, On a problem of Schiitte and ErdSs, Mathematical Gazette, 49 (1965), 290-293.
Algebraic Curves in Finite Geometry
235
[103] G. Tallini, Sui q-archi di un piano lineare finito di caratteristica p = 2, Atti dell'Accademia Nazionale dei Lincei, Classe di Scienze Fisiche Matematiche e Naturali, Rendiconti, 23 (1957), 242-245.
[104] M. Tallini Scafati, Archi completi in un 52,9 con q pari, Atti della Accademia Nazionale dei Lincei, Classe di Scienze Fisiche Matematiche e Naturali, Rendiconti, 37 (1964), 48-51.
[105] J. A. Thas, Some results concerning {(q + 1)(n - 1),n}-arcs and {(q + 1)(n - 1) + 1, n}-arcs in finite projective planes of order q, Journal of Combinatorial Theory, Series A, 19 (1975), 228-232. [106] J. A. Thas, Elementary proofs of two fundamental theorems of B. Segre without using the Hasse-Weil theorem, Journal of Combinatorial Theory, Series A, 34 (1983), 381-384.
[107] J. A. Thas, Complete arcs and algebraic curves in PG(2, q), Journal of Algebra, 106 (1987), 451-464.
[108] J. A. Thas, A combinatorial characterization of Hermitian curves, Journal of Algebraic Combinatorics, 1 (1992), 97-102.
[109] J. A. Thas, Projective geometry over a finite field, Chapter 7, in Handbook of Incidence Geometry (ed. F. Buekenhout), North-Holland, Amsterdam (1995), pp. 295-347.
[110] A. Thomason, Random graphs, strongly regular graphs and pseudorandom graphs, in Surveys in Combinatorics 1987 (ed. C. Whitehead), London Mathematical Society Lecture Note Series, 123, Cambridge University Press, Cambridge (1987), pp. 173-195.
[111] E. Ughi, On (k, n)-fold blocking sets which can be obtained as a union of conics, Geometriae Dedicata, 26 (1988), 241-246. [112] V. H. Vu, A strongly regular N-full graph of small order, Combinatorica, 16 (1996), 295-299.
[113] J. F. Voloch, On the completeness of certain plane arcs, European Journal of Combinatorics, 8 (1987), 453-456. [114] J. F. Voloch, Arcs in projective planes over prime fields, Journal of Geometry, 38 (1990), 198-200.
[115] J. F. Voloch, On the completeness of certain plane arcs II, European Journal of Combinatorics, 11 (1990), 491-496.
236
T. Szonyi
[116] J. F. Voloch, Complete arcs in Galois planes of non-square order, in Advances in Finite Geometries and Designs (eds. J. W. P. Hirschfeld, D. R. Hughes & J. A. Thas), Oxford University Press, Oxford (1991), pp. 401-406. [117] W. Waterhouse, Abelian varieties over finite fields, Annales Scientifiques de l'Ecole Normale Superieure, 2 (1969), 521-560.
[118] A. Weil, Sur les Courbes Algebriques et les Varietes qui s'en deduisent, Actualites Scientifiques et Industrielles, 1041, Hermann & Cie, Paris (1948).
[119] A. Weil, On some exponential sums, Proceedings of the National Academy of Science, 34 (1948), 204-207.
[120] F. Wettl, On the nuclei of a finite projective plane, Journal of Geometry, 30 (1987), 157-163.
[121] B. J. Wilson, Incompleteness of (nq + n - q - 2, n)-arcs in finite projective planes of even order, Mathematical Proceedings of the Cambridge Philosophical Society, 91 (1982), 1-8.
[122] F. Zirilli, Su una classe di k-archi di un piano di Galois, Atti della Accademia Nazionale dei Lincei, Classe di Scienze Fisiche Matematiche e Naturali, Rendiconti, 54 (1973), 393-397.
Department of Computer Science, Eotvos Lorand University, Muzeum krt. 6-8, H-1088 Budapest, Hungary and Department of Geometry, Jozsef Attila University, Aradi vertanuk tere 1, H-6720 Szeged, Hungary
[email protected] New Perspectives on Interval Orders and Interval Graphs William T. Trotter Summary Interval orders and interval graphs are particularly natural examples of two widely studied classes of discrete structures: partially ordered sets and undirected graphs. So it is not surprising that researchers in such diverse fields as mathematics, computer science, engineering and the social sciences have investigated structural, algorithmic, enumerative, combinatorial, extremal and even experimental problems associated with them. In this article, we survey recent work on interval orders and interval graphs, including research on on-line coloring, dimension estimates, fractional parameters, balancing pairs, hamiltonian paths, ramsey theory, extremal problems and tolerance orders. We provide an outline of the arguments for many of these results, especially those which seem to have a wide range of potential applications. Also, we provide short proofs of some of the more classical results
on interval orders and interval graphs. Our goal is to provide fresh insights into the current status of research in this area while suggesting new perspectives and directions for the future. 1
Introduction
A complex process (manufacturing computer chips, for example) is often broken into a series of tasks, each with a specified starting and ending time. Task A precedes Task B if A ends before B begins. When A precedes B, the output of A can safely be used as input to B, and resources dedicated to the completion of A, such as machines or personnel, can now be applied to B. When A and B have overlapping time periods, they may be viewed as conflicting tasks, in the sense that they compete for limited resources. This short paragraph is intended to motivate the formal definition of two of the most widely studied classes of discrete structures in all of combinatorial mathematics: interval orders and interval graphs. The main point to the discussion is that interval orders and interval graphs are important from an applications standpoint. This much is inescapable. They occur so naturally and with such frequency that they must be studied. Fortunately, the study of interval orders and interval graphs has yielded work of intrinsic interest and beauty, work that can be appreciated for its elegance independent of the fact that many find it useful and important. The remainder of this section includes a brief summary of the notation and terminology necessary for the balance of the paper. For a more comprehensive treatment of background material, the reader is referred to Peter Fishburn's monograph Interval Orders and Interval Graphs [36]. Other recommended sources for background information are the author's survey articles [115], [116], [120], [121] and monograph [118] and the books by Golumbic [48] and 237
William T. Trotter
238
Roberts [98].
Throughout this paper, we consider a partially ordered set (or poset) P = (X, P) as a structure consisting of a set X and a reflexive, antisymmetric and transitive binary relation P on X. We call X the ground set of the poset P, and we call P a partial order on X. The notations x < y in P, y > x in P and (x, y) E P are used interchangeably, and the reference to the partial order P is often dropped when its definition is fixed throughout the discussion. We write
x < y in P and y > x in P when x < y in P and x 54 y. When x,yEX, (x, y) P and (y, x) P, we say x and y are incomparable and write xIIy in P. When P = (X, P) is a poset, we call the partial order pd = {(y, x) : (x, y) E P} the dual of P and we let pd = (X, Pd). When P is a binary relation on X and Y C X, we denote the restriction
of P to Y by P(Y). When P is a partial order on X, Q = P(Y) is a partial order on Y and Q = (Y, Q) is called a subposet of P = (X, P). Also, we call Q the subposet determined by Y. When X1, X2,. . . , X,. C_ X, we will find it convenient to denote the subposets they determine by X1, X2 .... X,., respectively. In this article, we tend not to distinguish between isomorphic posets, so we abuse language slightly and say that a poset Q is contained in another poset P when Q is isomorphic to a subposet of P. Although we are concerned primarily with finite posets, i.e., those posets with finite ground sets, we find it convenient to use the familiar notation R, Q, Z and N to denote respectively the reals, rationals, integers and positive integers equipped with the usual orders. Note that these four infinite posets are total orders; in each case, any two distinct points are comparable. Total orders are also called linear orders, or chains. When X = X1 U X2 U U Xt is a partition and Li is a linear order on X2 for each i = 1, 2, ... , r, we let L = L1 < L2 < . . < L, denote the linear order on X defined by x < y in L if and only if x E Xi, y E Xj and either i < j or both i = j and x < y in Li.
For a positive integer n, we let n denote the n-element chain 0 < 1 < < n - 1. Somewhat inconsistently, we let [n] denote the n-element set { 1, 2, ... , n }. Also, when X is a set, we let (n) denote the set of all n-element subsets of X.
Let P = (X, P) be a poset, and let F
Qom)
: x E X} be a
family of posets indexed by the elements of X. Define the lexicographic sum of F over P, denoted E.EP J , as the poset Q = (Y, Q) where Y = {(x, y) : x E X, y E Y} and (x1, y1) < (x2, y2) in Q if and only if x1 < x2 in P, or if both x1 = x2 and yl < Y2 in Q2 . With this definition, a disjoint sum is just a lexicographic sum over a two-element antichain. In the remainder of this article, we will assume some familiarity with the basic concepts for partially ordered sets. The author's survey article on partially ordered sets [120] provides a thorough overview of the combinatorial aspects. Other sources for background material on posets are Brightwell's survey article [17] and the author's other survey articles [115], [112], [117] and [121].
Interval Orders and Interval Graphs
2
239
Interval orders and interval graphs A poset P = (X, P) is called an interval order if there is a function I
assigning to each element x E X a closed interval I (x) = [ax, bx] of a linearly ordered set L (usually, we take L as the real line lI8) so that for all x, y E X, x < y in P if and only if ay in L. We call I an interval representation
of P, or just a representation for short. For brevity, whenever we say that I is a representation of an interval order P = (X, P), we will use the alternate notation [ate, bx] for the closed interval I (x). Also, we let JI (x) I denote the length of the interval, i.e., II(x)I = bx - ay.
Note that end points of intervals used in a representation need not be distinct. In fact, distinct points x and y from X may satisfy I(x) = I(y). We even allow degenerate intervals. On the other hand, a representation is said to be distinguishing if all intervals are non-degenerate and all end points are distinct. It is easy to see that every interval order has a distinguishing representation. In fact, since we are concerned only with finite posets, we could have just as well required that all intervals used in the representation be open.
Analogously, a graph G = (V, E) is an interval graph when there is a function I which assigns to each vertex x E V a closed interval I (x) = [ay, b.'] from a linearly ordered set L so that {x, y} E E if and only if I(x) fl I(y) # 0. As before, we call I an interval representation of G and note that, if desired, we may assume I is distinguishing. Throughout this article, we will move back and forth between posets and graphs in discussions about a family of intervals. The interval graph determined by a family of intervals is just the incomparability graph of the interval order. Chains correspond to independent sets and antichains correspond to cliques.
3
Classical representation theorems
A good fraction of the early research on interval graphs and interval orders was focused on characterization issues. Recall that a graph is triangulated if it does not contain a cycle on four or more vertices as an induced subgraph. Also, a vertex x in a graph G is simplicial if its neighborhood is a complete subgraph of G, so a graph is triangulated if and only if every induced subgraph has a simplicial vertex. Triangulated graphs are a well studied class of perfect graphs (see [58] and Chapter 4 of Golumbic's monograph [48], for example). Obviously, interval graphs are triangulated, but it is natural to ask whether all triangulated graphs are interval graphs. This is not true. In fact, not all trees are interval graphs, e.g. the subdivision of K(1, 3) is not an interval graph. Three distinct vertices x, y and z in a graph G are said to form an asteroidal triple when for each two vertices in {x, y, z}, there is a path joining them, with no vertex on the path adjacent to the third. For example, the three leaves in a
William T. Trotter
240
subdivision of K(1, 3) form an asteroidal triple. In [83], Lekkerkerker and Boland proved that a triangulated graph is an interval graph if and only if it does not contain any asteroidal triples. They used this characterization theorem to provide a minimum list of forbidden subgraphs for interval graphs. This list includes the cycles on four or more vertices, three other infinite families and two isolated examples. One of these is the subdivision of K(1, 3). Other characterizations of interval graphs in terms of forbidden substructures have been provided by Gilmore and Hoffman [47] and by Ghouila-Houri [46]. Characterizations of interval graphs by forbidden subgraphs or forbidden substructures provide important structural information about the properties of interval graphs but do not necessarily yield a useful algorithm. Using a special kind of data structure called a PQ-tree, Booth and Lueker [15] produced an 0(n2) algorithm for testing whether a graph G on n vertices is an interval graph and producing the representation when it is. Characterization problems for interval graphs are closely related to characterization problems for comparability graphs. The classic paper of Gallai [45] provides a forbidden subgraph (again, in terms of induced subgraphs) characterization of comparability graphs with a minimum list including eight in-
finite families and 10 isolated examples. A comparability graph may have many different transitive orientations, but Gallai shows that if Tl and T2 are transitive orientations of the same comparability graph, then Tl may be transformed into T2 by a finite sequence of reversals applied to autonomous sets. Gallai's paper remains one of the deepest and most important contributions to this subject. Next, we discuss three important representation theorems which are essential to understanding the material which follows. First, a finite poset P = (X, P) is called a weak order if there exists a function f : X -* 1l8 so that for all x, y E X with x y,
1. x 1. Set Xj = Xj_1 - Aj_1. If Xj 0, let yj be the unique element of Xj for which the right end point rj = bye is minimum. Then let A. _ {x E Xj : byi E I(x)}. When the algorithm halts, we have a partition X = Al U A2 U ... U Ah into h antichains, and we have a chain C = {y1, y2, ... , yh} of cardinality h. Furthermore, every interval in the representation intersects the right end point of at least one interval in C. We call C the lexicographically least maximum chain of P, and we call the associated partition into antichains the canonical minimum partition. 5
Linear extensions and dimension When P and Q are binary relations on a set X, we say Q is an extension
of P when P C Q; a linear order L on X is called a linear extension of a partial order P on X when P C L. A set R of linear extensions of P is called a realizer of P when p = n R, i.e., for all x, y in X, x < y in P if and only if x < y in L, for every L E R. The minimum cardinality of a realizer of P is called the dimension of P and is denoted dim(P). Note that if P contains Q, then dim(Q) < dim(P). It is natural to ask what causes a poset to have large dimension. Here
is a partial answer. For integers n > 2 and k > 0, define the crown Sn as the poset of height two with n + k minimal elements a1, a2i ... , an+k, n + k maximal elements b1, b 2 ,- .. , bn+k and ordering ai < bj if and only if j E
{i + k + 1, i + k + 2,...,i - 1}. In this definition, the subscripts are interpreted cyclically, so that bn+k+l = b1i etc. When n > 3, the dimension of the crown Sn is given by the following formula [110]:
dim(Sn) = [2(n + 2 )1
(2)
For each k > 0, the poset SZ is the disjoint sum of k + 2 two-element chains,
so these posets have dimension 2. When n > 3, the crown Sn always has dimension at least 3. Posets in the family S = {S° : n > 2} are referred to as
Interval Orders and Interval Graphs
245
standard examples. Note that the dimension of S° is exactly n. Furthermore,
for each n > 3, So is n-irreducible, i.e., the removal of any point leaves a subposet having dimension n - 1. Also note that when n > 3, the standard example So is isomorphic to the family of 1-element and (n-1)-element subsets of {1, 2, ... , n } ordered by inclusion. The standard example S2 is somewhat of a special case. It has dimension two and is isomorphic to the disjoint sum of two 2-element chains, but it is not irreducible. We summarize some basic facts about dimension in the following proposition, referring the reader to [118] for proofs and references.
Proposition 5.1 Let P = (X, P) and Q = (Y, Q) be posets. Then: 1. dim(P + Q) = max{2, dim(e), dim(Q)}. 2. dim(P x Q)
dim(P) + dim(Q), with equality holding if P and Q have greatest and least elements.
3. The removal of a point from P decreases dim(P) by at most one.
4. If A is a maximum antichain in P, then dim(P) < JAI and dim(P) < max{2, IX - Al}.
5. dim(P) = dim(Pd). Note that the family of standard examples shows that inequalities 3 and 4 of Proposition 5.1 are best possible. We will also find it convenient to put inequality 1 in the preceding theorem in a more general setting. Here is the general formula for dimension and lexicographic sums (see [118]).
Proposition 5.2 Let P = (X, P) be a poset, and let.F = {Q5 = (Yr, P5) x E X1 be a family of posets. Then dim
(F)
= max{dim(e), max{dim(Q5) : x E X}}.
(3)
SEP
For additional background information on dimension, the reader is referred to the author's monograph [118], the survey article [63] on dimension by Kelly and Trotter and the survey articles [115] and [121]. The articles [112], [117] and [119] also discuss combinatorial problems for posets. Connections between dimension for posets and a wide range of combinatorial problems are discussed in [123], with greater detail provided in the monograph [118]. 6
Linear extensions of interval orders
When P = (X, P) is a poset, A, B C X and A fl B = 0, and L is a linear extension of P, we say B is over A in L when b > a in L, whenever a E A, b E B and allb in P. In applying this definition, it is important to note that we do not require that b > a in L, for all a E A and b E B, only the incomparable pairs. The following elementary result was first discovered by Rabinovitch [94].
William T. Trotter
246
Theorem 6.1 Let P = (X, P) be an interval poset, and let A, B C X with A fl B = 0. Then there exists a linear extension L of P with B over A in L. Proof Let I be a distinguishing interval representation of P. For each x E X, let px = a,, if x E A and ps = bx, otherwise. Then define a linear extension L by setting x < y in L if and only if ps < py in R. More generally, the following proposition, first noted by Felsner in [28], is an easy exercise.
Proposition 6.2 Let P = (X, P) be an interval order, and let I be any distinguishing interval representation of P. If L is a linear extension of P, then it is possible to choose for each x E X a point p,, E I(x) so that x < y in L if and only if p < py in R. 7
Dimension of interval orders
It is natural to ask whether an interval order can have large dimension. If the answer is yes, it cannot be due to the presence of large standard examples,
as no interval order contains any of them. Note that for each n > 2, the subposet of So determined by a1, a2, b1 and b2 is isomorphic to 2 + 2.
Nevertheless, interval orders may have large dimension, and to explain how this may occur, we introduce a standard example of an interval order. For an integer n > 2, let In = ((2v), P,) denote the interval order defined by the representation I({i, j}) = [i, j]. To avoid confusion with the family of standard examples discussed previously, we call the interval orders in the family {In : n > 2} canonical interval orders. The following result is due to Bogart, Rabinovitch and Trotter [10]. Theorem 7.1 For every integer t, there exists an integer no so that if n > no, then the dimension of the canonical interval order In is larger than t. Proof Evidently, dim(In) is a non-decreasing function of n. We assume that dim(In) < t, for all n > 2 and obtain a contradiction when n is sufficiently large in terms of t. Let i, j, and k be distinct integers with 1 < i < j < k < n. Then {i, j} 11 { j, k} in Pn, so if R = {L1, L2, ... , Lt} is a realizer of Pn, then we
may choose a E {1, 2, ... , t } so that {i, j J > f j, k} in L,,. This is a coloring of the 3-element subsets of 11, 2, ... , n } with t colors. If n is sufficiently large,
there exists a 4-element subset S = {i < j < k < l} and an integer a E 11, 2, ... , t } so that all 3-element subsets of S are mapped to a. This implies that {i, j} > {j, k} > {k, 1} > {i, j} in Lam, which is a contradiction.
Now that we know that interval orders can have large dimension, we pause to discuss some of the special properties interval orders exhibit.
Let P = (X, P) be a poset and let X = Xl U X2 be a partition of X into disjoint non-empty subsets. It is natural to ask whether one can say anything
Interval Orders and Interval Graphs
247
about the dimension of dim(P) given information about dim(Xi) and dim(X2)For posets in general, the answer is no. For example, for each n > 2, consider the partition of the point set of the standard example S° into minimal elements and maximal elements. The dimension of Sn is n but the two antichains have dimension 2. For interval orders, things are different. The next result follows easily from Proposition 6.2.
Lemma 7.2 Let P = (X, P) be an interval order, and let X = Xl U X2 be a partition of X into disjoint non-empty subsets. If L1 and L2 are linear extensions of the subposets P1 and P2 induced by X1 and X2 respectively, then there exists a linear extension L of P so that L1 = L(X1) and L2 = L(X2).
Theorem 7.3 Let P = (X, P) be an interval order, and let X = X1 U X2 be a partition of X into disjoint non-empty subsets. Then dim(P) < 2 + max{dim(Xi), dim(X2)}.
(4)
Proof Let t = max{dim(Xi), dim(X2)}. From Lemma 7.2, we know that there exists a family F = {L1, L2, ... , Lt} of linear extensions of P so that .Fi = {L1(Xi), L2(Xi), ... , Lt(Xi)} is a realizer of Xi, for i = 1, 2. Then let M1 and M2 be linear extensions of P so that X1 is over X2 in M1 and X2 is over X1 in M2. It follows that {M1i M2} UT is a realizer of P. When one of the two sets in the partition is the set of maximal elements, we can do a little better. We leave the proof as an exercise.
Theorem 7.4 Let P = (X, P) be an interval order which is not an antichain. If X1 is the set of all maximal elements, and X2 = X - X1, then dim(P) < 1 + dim(X2).
(5)
Now it is natural to ask whether we can say anything about what must be contained in an interval order of large dimension. Here we present a partial answer. In Section 10, we will give a more complete answer. For now, we are content to show that an interval order of large dimension must contain a long chain.
Theorem 7.5 An interval order of height h has dimension at most h + 1. Proof Let P = (X, P) be an interval order of height h, let I be a distinguishing representation of P and let X = Al U A2 U ... U Ah be the canonical partition into antichains. Note that if x < y in P, x E Ai and y E A;, then i < j. For each i E [h], let Li be a linear extension of P with Ai over X - Ai in Li. Then let Lh+1 = Ld(A1) < Ld(A2) < . . . < Ld(Ah)
William T. Trotter
248
Although interval orders can have large dimension, this is not true for semi-orders. The following result is due to Rabinovitch [93].
Theorem 7.6 If P = (X, P) is a semi-order, then dim(P) < 3. Proof Let P = (X, P) be a semi-order, let I be a distinguishing representation of P and let let X = A,UA2U. . UAh, be the canonical partition into antichains.
Let 0 = U{Aj : 1 < j < h, j odd} and £ = lJ{Aj : 1 < j < h, j even}. Let Ll and L2 be linear extensions of P with £ over 0 in Ll and 0 over £ in L2. Then let L3 = Ld(A,) < Ld(A2)
1, there exists an integer Sk so that if P = (X, P) is an interval order in which the maximum size of a chain C for which there exists a point x incomparable to all points of C is at most k, then dim(P) < sk. 8
Critical pairs and alternating cycles
In arguments to follow, we will find it convenient to take advantage of a technical detail in the proof of Theorem 7.6. Let L be an arbitrary linear order on X. Define linear extensions Ld and L,,, of P as follows. Set x < y in Ld if and only if one of the following conditions holds:
1. D(x) C D(y). 2. D(x) = D(y) and U(y) C U(x).
3. D(x) = D(y), U(y) = U(x), and x < y in L. Dually, set x < y in L. if and only if one of the following conditions holds:
1. U(y) C U(x).
2. U(y) = U(x) and D(x) C D(y). 3. U(y) = U(x), D(x) = U(x), and x > y in L.
is a Now let F be a family of linear extensions of P. Then {Ld, L.} U realizer of P if and only if for every x, y E X with x I I y in P, D(x) C D(y), and U(y) C U(x), there exists L E F with x > y in L. This last observation is a special case of a somewhat more general situation.
For an arbitrary poset P = (X, P), let inc(P) = {(x, y) E X X X : x1ly in P}. Then a family 1Z of linear extensions of P is a realizer of P if and only if for every (x, y) E inc(P), there exists L E R so that x > y in L. Call a pair
Interval Orders and Interval Graphs
249
(x, y) E inc(P) a critical pair if u < x in P implies u < y in P and v > y in P implies v > x in P, for all u, v E X. Then let crit(P) denote the set of all critical pairs. It follows that 1Z is a realizer of P if and only if for every (x, y) E crit(P), there exists some L E 1Z so that x > yin L. We say L reverses the incomparable pair (x, y) when x > y in L. Let S C inc(P). We say that L reverses S when x > yin L, for every (x, y) E S. For an integer k > 2, a subset S = {(xi, yi) : 1 < i < k} C inc(P) is called an alternating cycle when xi _< yi+1 in P, for all i = 1 , 2, ... , k. In this last definition, the subscripts are interpreted cyclically, i.e., Yk+1 = yi. An alternating cycle S = {(xi, yi) : 1 < i < k} is strict if xi < yj in P if and only if j = i + 1, for all i, j = 1, 2, ... , k. When an alternating cycle is strict, the following three statements hold: 1. The elements in {x1, x2, ... , xk} form a k-element antichain. 2. The elements in {yi, Y 27 .. , AI form a k-element antichain.
3. If i, j E [k] and xi > yj, then j = i + 1 and xi = yp The following elementary result is due to Trotter and Moore [124]. See [118]
for a short proof and a number of applications.
Theorem 8.1 Let P = (X, P) be a poset and let S C inc(P). Then the following statements are equivalent.
1. There exists a linear extension L of P which reverses S.
2. S does not contain an alternating cycle.
3. S does not contain a strict alternating cycle. 9
Interval orders and shift graphs
Although it has been known for many years that interval orders of large height must contain long chains, it has only been in the last few years that relatively tight bounds have been found. The relationship between height and dimension in interval orders is best explained via a connection with a graph coloring problem. Fix integers n and k with 1 < k < it. We call an ordered pair (A, B) of k-element sets a (k, n)-shift pair if there exists a (k + 1)-element < ik+1} C {1, 2, ... , n } so that A = {il, i2i ... , ik} subset C = {i1 < i2 < and B = {i2i i3, ... , ik+1}. We then define the (k, n)-shift graph S(k, n) as the graph whose vertex set consists of all k-element subsets of 11, 2, ... , n } with a k-element set A adjacent to a k-element set B exactly when (A, B) is a (k, n)shift pair. The shift graph S(1, n) is just a complete graph on it vertices. It is customary to call a (2, n)-shift graph just a shift graph; similarly, a (3, n)-shift graph is called a double shift graph.
William T. Trotter
250
One of the folklore results of graph theory is the following formula for the chromatic number of the shift graph (throughout this paper, we use the notation lgn to denote the base 2 logarithm of n). Theorem 9.1 The chromatic number of the shift graph S(2, n) is Ilg n] . The proof of Theorem 7.1 establishes the following lower bound.
Proposition 9.2 The dimension of the canonical interval order In is at least as large as the chromatic number of the double shift graph S(3, n). In turn, the next result relates the determination of the dimension of the family of canonical interval orders to the classical enumeration problem known as Dedekind's problem: estimate the number of antichains in the poset 2t, the cartesian product of t two-element chains. This poset is just the subset lattice,
the family of all subsets of [t] partially ordered by inclusion. The next four results are due to Fiiredi, Hajnal, Rodl and Trotter [44]. Theorem 9.3 The chromatic number of the double shift graph S(3, n) is the least positive integer t for which there are at least n antichains in the subset lattice 2t.
Proof Suppose that there are n antichains in the subset lattice 2t. We show that the chromatic number of S(3, n) < t. Let Q be the partial order defined on the antichains of 2t by setting A < 13 in Q if and only if for every S E A, there exists B E 13 so that A C _. B. Then let L be any linear extension of Q, < An in L. For each i, j with 1 < i < j < n, and suppose that Al < A2 < let B(i, j) E A; be a set so that there is no set A E Ai with A C B(i, j). Then for each i, j, k with 1 < i < j < k < n, choose an element a E B(j, k) -B(i, j), and set ¢({i, j, k}) = a. Then 0 is a coloring of S(3, n). Conversely, if the chromatic number of S(3, n) is at most t and 0: (3]) -p [t]
is a coloring, we define for each i, j with 1 < i < j
n, the set B(i, j) _
{¢({i, j, k}) : j < k < n}. Then for each i E [n], set 13i = {B(i, j) : i < j < n}. Partial order each Bi by inclusion and let Ai be the maximal elements. Then each Ai is an antichain in 2t and Ail # Ail when it 0 i2. Although no closed form solution to Dedekind's problem has been found, relatively tight estimates are known (see [77], e.g.). For our purposes, we may subsets of size It/2]. use the estimate which results as follows. There are Any subset of these sets forms an antichain in 2t.
Theorem 9.4 The chromatic number of the double shift graph satisfies:
x(S(3,n))=lglgn+(1/2+o(1))lglglgn.
(6)
Interval Orders and Interval Graphs
251
Hopefully, the reader has noticed the following subtlety to the dimension problem for canonical interval orders. The lower bound depends heavily on the use of repeated end points, a phenomenon which we can eliminate by modifying the representation. After the modification, it is conceivable that the dimension problem is much harder. However, this turns out not to be the case.
Theorem 9.5 Let n and t be positive integers with n > 2. If
n j2. Dually, set [il, ill < [i2i j2] in M2 if jl < j2, or if both jl = j2 and it > i2. It remains to find t + 1 additional linear extensions L1, L 2 ,-- . , Lt+l so that when i - 1 < i2 < jl < j2i there is at least one La so that [il, ill > [i2,j2] in La. Let s = and let S1, S2, ... , Ss be a listing of all the It/2]-element subsets of [t]. Note that A = IS,, S2i... , S} is an antichain in the subset lattice 2t. Also let X denote the set of all 0-1 sequences of length s, i.e., the elements of X are functions from [s] to 10, 11. We let L be the lexicographic order on X. By this, we mean that if f,g E X and j is the least integer for which f (j) g (j ), then f < g in L if and only if f (j) < g (j) . This implies f (j) = 0 and g(j) = 1. Now let L' _ [fl, f2, ... , fn] be the restriction of L to n distinct elements of X. Now let [i1, jl] and [i2,j2] be elements of In with it < i2 < it < j2. Note that we allow the possibility that i2 = jl. Set E = {il, j2, jl, j2} so that E has either 3 or 4 elements. Choose the least integer k1 so that I{ fs(k1) : i E E}I > 1. Note that f1 (kl) = 0 and fj2(k1) = 1. Furthermore, exactly one of the following statements holds:
1. For every ieEwith i0ii, Mil) =1. 2. For every i E E with i # j2i fz(jl) = 0. 3. fz2(k1) = 0 and ffj,(k1) = 1.
When the third of these statements holds, we must have JEl = 4, but the first two may occur with either JEl = 3 or IEI = 4. Also, when the third statement holds, we will require that [il, ill > [i2, j2] in Lt+1. When the first statement holds, set E' = E - {i1}, and when the second statement holds, set E' = E - {j2}. In either case, let k2 be the least element where I { fz(kl) : i E E'}l > 1. If the first statement holds, choose a E Sk1- Sk2 and require [il, jl] > [i2, j2] in La. If the second statement holds, choose a E Sk2 - Sk1 and require [i1, jl] > [i2, j2] in La. We leave it as an exercise that such linear extensions exist.
William T. Trotter
252
The preceding theorem shows that the lower bound provided in inequal-
ity (6) is also an upper bound. With a little more work, the same kind of estimate works for arbitrary interval orders (see [44] for details).
Theorem 9.6 The maximum dimension d(h) of an interval order of height h satisfies:
d(h)=lglgh+(1/2+o(1))lglglgh.
(8)
Before closing this section, we comment that the dimension problem for interval orders is closely related to the problem of determining the dimension of the poset consisting of all 1-element and 2-element subsets of {1, 2, ... , n}, partially ordered by inclusion. Spencer [107) was the first to establish the connection between this problem and the classic result of Erdos and Szekeres concerning monotonic subsequences of a sequence of integers. In recent years, there has been rapid progress in estimating the dimension of posets consisting of layers of the subset lattice. A summary of this work together with additional references is provided by Trotter in [121]. 10
Interval orders and overlap graphs
A graph G = (V, E) is called an overlap graph when there exists a function I assigning to each vertex x E V a closed interval I (x) _ [ax, b'] of R so
that for all x, y E V, {x, y} E E if and only if I(x) flI(y) 0, I(x) Z I(y) and I (y) Z I (x), i.e., the intervals intersect, but neither is contained in the other. Again, we call the function I a representation of the overlap graph G. If required, we may assume that a representation of an overlap graph is distinguishing. In general, overlap graphs need not be perfect, e.g. a cycle on 5 vertices is
an overlap graph. However, when all the intervals used in the representation intersect, then the graph is perfect, and it is easy to color such graphs.
Proposition 10.1 Let I be a distinguishing representation for an overlap graph G = (V, E). If I(x) fl I(y) 0, for all x, y E V, then G is the comparability graph of a poset P = (V, P) with dim(P) < 2, so G is perfect. Furthermore, the First Fit algorithm will provide an optimal coloring of G if the vertices are colored in the order determined by the left end points.
Proof Let Ll and L2 be the linear orders on V determined by left and right end points, respectively. Then let P = Ll fl L2. Clearly, {x, y} is an edge of G if and only if x and y are comparable in the the poset P = (V, P). From its definition, we know that dim(P) < 2. The First Fit algorithm applied to the vertices in the order of their left end points is the minimum antichain partition described in the proof of Theorem 4.1. When the intervals used in the representation do not share a common point, it is not immediately clear that there is any bound on the chromatic number of
Interval Orders and Interval Graphs
253
an overlap graph in terms of the maximum clique size. The first proof of this fact is due to Gyarfas [55] and the best bounds to date are due to Kostochka and Kratochvil [78]. Recall that the notation f (k) = Sl(g(k)) means that there exists a positive constant c and an integer ko so that f (k) > cg(k), for
allk>ko. Theorem 10.2 Let m(k) = max{x(G) : G is an overlap graph with maximum clique size k}. Then 1. m(k) =12 (k log k). 2. m(k) = 0(2 k).
Quite recently, the concepts used in the proof of Theorem 10.2 have been
applied by Kierstead and Trotter [75] to solve a long standing problem in dimension theory for interval orders. We outline this work, but we do not aim for the best possible constants.
Theorem 10.3 For every interval order Q = (Y, Q), there exists an integer to so that if P = (X, P) is any interval order with dim(P) > to, then P contains a subposet isomorphic to Q. Since every interval order Q is isomorphic to a subposet of the canonical interval order I,1, provided n is sufficiently large, Theorem 10.3 is equivalent to
showing that for every integer n > 2, there exists an integer t,, so that if P is an interval order with dim(P) > t, then P contains a subposet isomorphic to the canonical interval order I,,.
Let P = (X, P) be an interval order. For an integer m > 2, we call a subposet T of P an m-tower T when 1. T contains an m-element chain Z = {zl < z2 < . . . < zm}, and
2. For every pair i, j with 1 < i < j < m, T contains an element w(i, j) which is incomparable with zi, z1,. .. , zj and comparable with all other elements of Z.
It is an easy exercise to show that if P contains a 3n-tower, it contains a subposet isomorphic to I,,. So, Theorem 10.3 is also equivalent to the following somewhat more technical result.
Theorem 10.4 For every integer m > 2, there exists an integer tm so that if P is an interval order with dim(P) > tm, then P contains a subposet isomorphic to an m-tower.
Proof We proceed by induction on m. An interval order which does not contain a 2-tower is a weak order and has dimension at most 2. So it suffices to take t2 = 3. Now consider a value of m > 3 and assume that there exists an integer tm_1 so that any interval order whose dimension is at least tm_1
William T. Trotter
254
contains an (m - 1)-tower. Now let P = (X, P) be any interval order whose dimension t is at least tm_1 + 9. We show that P contains an m-tower. The key idea in the remainder of the proof is the notion of distance in the overlap graph. Let I be a distinguishing representation of P = (X, P). We proceed to build a realizer of P, starting with the two linear extensions M1 and M2, the orderings determined by the left and right end points respectively in the representation I. The important thing to notice is that it only remains to reverse critical pairs of the form (x, y), where ax < ay < b, < by. In particular, x and y are adjacent in the overlap graph. Let G denote the overlap graph determined by I, and let G1, G2, ... , Gs denote the components of G. Then let Xi = (Xi, Pi) be the subposet determined by the vertex set of Gi. For each i E [s], define the root of Gi to be the unique vertex in Gi whose left end point is minimal. We denote the root of Gi
by ri. For every vertex x E Gi, let d(x, ri) denote the distance from x to ri in Gi. The following key lemma is due to Gyarfas [55]. We leave the proof as an exercise.
Lemma 10.5 Let i E [s] and let j > 0. Then let x, y and z each be at distance j from the root ri of a component Gi of G. If i(z) C i(x) fl I (y), and ri = uo, ui,... , ui = z is a shortest path from ri to z in H, then I (ui_1) I(x) U I(y). Next we classify all vertices of G as either left or right, and we denote the set of all left vertices by £ and the set of all right vertices by R. A vertex x belongs to G if and only if there exists a shortest path ri = uo, u1, . . . , ui = x from the root of the component to which x belongs to x so that the left end point of I(ui_1) is less than the left end point of I(x). Then set R = X - L. Similary, we classify all vertices as either even or odd and denote these two sets by £ and 0, respectively. A vertex x belongs to £ if and only if the distance from x to the root of the component to which it belongs is even. Then
set O=X-S.
Then let M3, M4, M5 and M6 be linear extensions of P with
1. £ over R in M3i 2. R over £ in M4;
3. £ over 0 in M5; and 4. 0 over £ in M5. It follows that we may assume that there is a pair i, j with i E [s] and j > 2 so that the subposet Q = (Y, Q) determined by all left vertices at distance j from the root ri of component Gi has dimension at least tm_1 + 3. Consider the following recursive definition. Set Yo = Y. If Yk has already
been defined for some k > 0 and the dimension of the subposet Yk is less
Interval Orders and Interval Graphs
255
than t,,,,_1 + 1, set Zk+1 = Yk and halt. If on the other hand, the dimension of Yk is at least tm_1 + 1, let yk+1 be the unique element of Yk with dim(W(yk+l, Y)) > tm-i + 1 whose left end point is as small as possible. Then set Bk+l = {y E Yk+1 byk+l E I(y)J, Zk+1 = W (yk+l, Y k) - Bk+l and Y+1 = Y - W (yk+i, Yk). It follows that dim(Yi+l) = tm_1 + 1 and dim(Zi+1) = tm-1 Suppose this recursive definition halts in a partition Y = Z U B, where Z = Zl U Z2 U Z, and B = Bl U B2 U Bs_1. Then dim(Z) = tm_1i so that dim(B) > tm_1 + 1. Also, note that f o r each i = 1, 2, ... , s - 1, the inductive hypothesis implies that Zi contains an (m - 1)-tower. :
Since the dimension of B is large, it follows (being very generous) that there are integers kl and k2 with 1 < kl < k2 - 3 and elements bk, E Bk,, bk2 E Bk2 so that bk, I I bkz in P. It follows that the interval for bk, properly contains intervals
from two disjoint (m - 1)-towers, one from Zk,+1 and the other from Zk,+2. Choose a vertex x from Zk,+l, and consider the (m - 1)-tower T from Zk2+2. For each vertex, y E T, the interval corresponding to the vertex just before y
on the shortest path from ri to y properly overlaps the interval for y. By the lemma, this interval also has a left end point which precedes the left end point of bk,. Thus this interval also intersects x. It follows that P contains an m-tower. 11
Semi-orders and balancing pairs Let P = (X, P) be a poset and let F = {M,. .. , Mt} be a multiset of
linear extensions of P. Consider the linear extensions of F as outcomes in a uniform sample space. For a distinct pair x, y E X, the probability that x is over y in F, denoted Prob,F[x > y], is defined by
ProbF[x>y]=1I{i:1yinMi}l.
(9)
In this section, we are concerned with the family A(P) consisting of all linear extensions of P. We let A(P) = JA(P)E. For this family, we drop the subscript and just write Prob[x > y]. Note that Prob[x > y] = 0, when x < y
in P; Prob[x > y] = 1, when x > y in P and 0 < Prob[x > y] < 1, when x1Iy in P. In 1969, S. S. Kislitsyn [76] made the following conjecture, which remains one of the most intriguing problems in the combinatorial theory of posets.
Conjecture 11.1 If P = (X, P) is a finite poset which is not a chain, then there exists an incomparable pair x, y E X so that
1/3 < Prob(x > y) < 2/3.
(10)
This conjecture was made independently by both M. Fredman and N. Linial, and many papers on this subject attribute the conjecture to them. It is now
William T. Trotter
256
known as the 1/3-2/3 conjecture. If true, the conjecture would be best possible, as shown by 2 + 1. The first major breakthrough in this area came in 1984, when Kahn and Saks [62] used the Alexandrov/Fenchel inequalities for mixed volumes to prove the following result.
Theorem 11.2 If P = (X, P) is a finite poset which is not a chain, then there exists an incomparable pair x, y E X so that 11 < Prob[x > y]
y] < 5
i
(12)
As pointed out in [18], there is an infinite semi-order for which the inequality in Theorem 11.3 is best possible, so that the 1/3-2/3 conjecture is false if one attempts to extend it to infinite posets. However, for finite semi-orders, we can do even better. For a poset P = (X, P), we say x covers y and write
x :> y in P when x > y in P and if x > z > y in P, then either x = z or y = z. The next result is due to Brightwell [16].
Theorem 11.4 If P = (X, P) is a finite semi-order which is not a chain, then there exists an incomparable pair x, y E X so that
3 < Prob[x > y] < 3
(13)
Proof Suppose that the theorem is false. Choose a counterexample P = (X, P) with I X I = n minimum. Then let I be a distinguishing representation. Label the points of X as x1, x2, ... , xn in the order determined by left end points. Define a linear order L on P by setting x < y in L if and only
if Prob[x > y] < 1/3. Clearly, L is a linear extension of P. Furthermore L orders X as x1 < x2 < < xn. We claim that xi II xi+1, for all i = 1, 2, ... , n - 1. To the contrary, suppose xi < xi+1 in P. Then P is the lexicographic sum over a two-element chain of the subposets determined by {x1, x2i ... , xi} and {xi+1, x 2 , . .. , xn}. One of these posets is not a chain, and we immediately contradict our choice of P = (X, P) as a minimum counterexample. We say that a point xj separates xi and xi+1 from above if xj :> xi and xj Ilxi+1 in P. Dually, we say xj separates xi and xi+1 from below if xi+1 :> xj
Interval Orders and Interval Graphs
257
and xillxi in P. Finally, we say xj separates xi and xi+1 if it separates them from above or from below. Note that if xj separates xi and xi+1 from above, then Xk < xj in P, for all k = 1, 2, ... , i. Dually, if Xk separates xi and xi+1 from below, then xj < xk in P, for all k = i+1, i+2,. .., n. So each xj separates at most two pairs, one from above and one from below. Furthermore, x1 and x2 do not separate pairs from below, while xn_1 and xn do not separate pairs from
above. It follows that there at most 2(n - 4) + 4 = 2n - 4 pairs (i, j) so that xj separates xi and xi+l. From this we conclude that there is an integer i (in fact, there are at least two such values) for which there is at most one integer j so that xj separates xi and xi+l. We show that 1/3 < Prob[xi > xi+1] < 2/3.
Let A(P) be the set of all linear extensions of P, and let JA(P) I = t. Set Al = {L E A : xi < xi+1 in L, but there is no element of X which separates xi and xi+1 between them in L}; A2 = {L E A - Al : xi < xi+1 in L}; and A3 = A - (A1 u A2). Then JA31/t = Prob[xi > xi+1] < 1/3. Consider the map h: Al --+ A3 defined as follows. For a linear extension L E A1, form h(L) by exchanging xi and xi+1 Clearly, the map h is an injection. It follows that JA11 < JA31. Furthermore, IA31/t = Prob[xi > xi+1] < 1/3, so IA2I > t/3. In particular, there exists a unique element xj which separates xi and xi+1. If xj separates from above, 1/3 < Prob[xi+l > xj] < 2/3. If xj separates from below, then 1/3 < Prob[xj > xi] < 2/3. There are some other special classes of posets for which the 1/3-2/3 conjecture is known to be true. For example, Fishburn, Gehrlein and Trotter [39] showed that it is valid for all posets of height 2.
In a poset P = (X, P), a sequence (x1i x2, ... , xn) of length n > 3 is called a linear extension majority cycle, or just an LEM cycle for short, when Prob[xi > xi+l] > 2, for all i E [n]. It is an easy exercise to show that semiorders do not contain LEM cycles, but Brightwell, Fishburn and Winkler [19]
show that LEM cycles can exist in interval orders-in fact, even in interval orders having dimension at most two. 12
Interval orders and extremal problems
Here are two interesting extremal problems involving semi-orders. The first problem is investigated by Fishburn and Trotter in [41]. For integers n and k with 0 < k < (2), let Q(n, k) denote the family of all posets with n points and k comparable pairs. Then set e(n, k) = max{IA(P)I : P = (X, P) E Q(n, k)}.
Theorem 12.1 Every poset P = (X, P) E Q(n, k) with JA(P) I = e(n, k) is a semi-order.
Proof Suppose that P = (X, P) E Q(n, k), JA(P) I = e(n, k), but that P is not a semi-order. Suppose further that P is not an interval order. Then P contains a subposet isomorphic to 2 + 2. Label the 4 points in the copy of 2 + 2 as {x, y, u, v}, so that u E D(x) - D(y) and v E D(y) - D(x). Of all copies of
William T. Trotter
258
2 + 2 in P, we may assume that we have chosen one so that 1U(x)1 + IU(y)I is minimum. It follows that one of U(x) and U(y) is a subset of the other. Without loss of generality, we assume that U(x) C U(y). Let P' = (X, P') be
the poset obtained from P by replacing the relations z < y by z < x for all z E D(y) - D(x). Then P' E Q(n, k). Interchanging the points x and y transforms a linear extension from A(P) A(P') into a linear extension from A(P') - A(P). Furthermore, this map is an injection. It is not a surjection, because any linear extension with y < u < v < x is not in the image of the map. The contradiction shows that P is an interval order. Now assume that P contains a subposet isomorphic to 3 + 1, and label the elements in the 3-element chain so that x < y < z in P. Label the element incomparable to these three points as w. Now form a poset P" = (X, P") E Q(n, k) by replacing relations t < y by t < w for all t E D(y) - D(w). Then P' E Q(n, k). As before, P" has more linear extensions than P. The second problem sounds similar. It was posed to me by Peter Winkler. Define the flexibility of a poset P = (X, P), denoted flex(P), by
flex(P) = E JU(x) + D(x) 12.
(14)
xEX
Then the same kind of argument used to prove Theorem 12.1 can be used to show that among all posets with n points and k comparable pairs, those with maximum flexibility are semi-orders. Despite our knowledge about the structure of the extremal posets, little progress has been made in solving either of these problems in full generality. Now here is an interesting extremal problem for posets on which significant results have been obtained for interval orders. When L is a linear extension of P = (X, P), let j(L, P) count the number of consecutive pairs of elements in L
which are incomparable in P. The jump number of P is then the minimum value of j (L, P) taken over all linear extensions of P. In [89], Mitas shows that determining the jump number of an interval order is NP-complete. However, Mitas [89], Felsner [29] and Syslo [108] have (independently) given a polynomial algorithm for approximating the jump number within a ratio of 3/2. Bogart and Stellpflug [11] define the representation length of a semi-order as the least positive integer k for which it has a a representation using intervals of length k with integer endpoints. For each k > 1, they provide a forbidden subposet characterization of semi-orders with representation length k. For interval orders, we have the following natural extremal problem posed
by Peter Fishburn in [36]. Given an interval order P, find the least positive integer k for which P has a representation using intervals having k distinct lengths. This parameter is called the interval count of P. Two interesting questions are immediate. First, what is the maximum value of the interval count of an interval order on n points? Second, can the removal of a single point drop the interval count by an arbitrarily large amount?
Interval Orders and Interval Graphs 13
259
Interval orders and hamiltonian paths Considered as a graph, the diagram of a poset of height h cannot have
chromatic number exceeding h. However, the "partite" construction of Nesetril and Rodl [90] shows that for every integer h, there exists a poset P of height h so that the chromatic number of the diagram of P is exactly h. For interval orders, the situation is completely different, and the chromatic
number of the diagram of an interval order of height h is much less than h. The open intervals with integer end points in {1, 2, ... , h+ 11 form an interval order of height h. Furthermore, the diagram of this interval order is just the shift graph S(2, h+ 1), a graph whose chromatic number is exactly [lg(h+1)]. Surprisingly, this is not far from best possible.
Let t be a positive integer, and let S = (So, Sl,... , Sh) be a sequence of sets. Felsner and Trotter [34] called F an a-sequence if S1 Z So and
S, - (Si U S2_1) # 0, for all i, j with 1 < i < j < h. Define a(t) to be the maximum h for which there exists an a-sequence (So, Si,.. ... , Sh), with each Si a subset of [t]. For example, a(3) = 5 as evidenced by the a-sequence A = (0, {1}, {2}, {3}, {1, 3}, {1, 2, 3}). Note that any subsequence of consecutive terms from an a-sequence is again an a-sequence. Let D(h) denote the maximum chromatic number of an interval order of height h. Clearly, D(1) = 1 and D(2) = 2.
Theorem 13.1 For each h > 2, D(h) is the least t for which a(t) > h. Proof We first show that if a(t) > h, then D(h) < t. Let S = (So, Si,... , Sh) be an a-sequence of subsets of [t], and let P be an interval order of height h. Then let I be a distinguishing representation of P. Let be the lexicographically least maximum chain be C = {yl < y2 < . . . < yh}, and let the canonical partition into antichains be X = Al U A2 U ... U Ah. For each i E [h], let ri = by:. Then let ro be any real number with ro < a,, in l[8, for every x E X.
For each x E X, set i(x) = max{i : 0 < i < h, ri < ay} and jx = max J j : 1 < j < h, rj E I (x) }. Note that ix < jx, for every x E X. We then define a coloring 0: X -+ [t] as follows. If ix = 0, choose 0(x) E Sj, - So. If iy > 1, choose O(x) E Sjr - (Six U Si.-,). We claim that 0 is a proper coloring of the diagram of P. Suppose that
x h. For integers h, m > 2, let P(h, m) denote the interval order determined by the family of all closed intervals with length at least m - 1 having integer end points from {1, 2, ... , m(h + 1) - 1 }. Note that the height of P(h, m) is h. We now show that for each h > 2, there exists an integer mo so that if m > mo and the chromatic number of the diagram of P(h, m) is t, then a(t) > h. In fact, we show that the choice mo = 2h2 works.
William T. Trotter
260
Fix h > 2 and then let m be any integer with m > mo. Suppose that the chromatic number of the diagram of P(h, m) is t. Note t < h. Now suppose that ¢ is a coloring of the diagram of P(m, h) using colors from [t]. For each
j=1,2,...,m(h+1)-1, let A;={¢([i,j]):1 y.
sequence (So, S1, .
Felsner and Trotter [34] conjecture that
a(t) = 2t-1 + I
t
2 11
.
(15)
If this conjecture is true, then an a-sequence S of subsets of [t] of maximum size has the following property. If we form a new sequence 9-l from S by inserting between two consecutive sets in S their union, when the first set is not a subset of the second, then we get a listing of all 2t subsets of [t]. For example, from the 6 term a-sequence of subsets of [3] given above, this listing is (0, {1}, {1, 2}, {2}, {2, 3}, {3}, {1, 3}, {1, 2, 3}). This listing is a special kind
of hamiltonian path in the t-cube. Whenever a set appears in the list, all of its subsets, with at most a single exception, appear previously. If there is an exception, it is listed next. We call such a path an order-preserving hamiltonian path in the t-cube. This is a slight abuse of the concept of order-preserving, but it is the strongest notion that makes sense. It is known that there are order-preserving hamiltonian paths in the t-cube for 1 < t < 8, but the general question is open. We should point out that attempts to settle whether equation (15) is always valid have produced the best known partial result on the well known "middle two levels" problem. The origins of the problem are a bit unclear, but it was first told to me by Ivan Havel during a visit to Prague.
Problem 13.2 Is the diagram of the poset consisting of all k-element and (k+l)-element subsets of a (2k+1)-element set, partially ordered by inclusion, a hamiltonian graph? We refer the reader to [34] and [99] for details.
Interval Orders and Interval Graphs
14
261
On-line and un-cooperative coloring
An on-line optimization problem, such as on-line graph coloring, can be considered as a two-person game involving a Builder and a Colorer. The game is played in a series of rounds with the players alternating turns. Each instance of on-line coloring involves two parameters: an integer t and a graph G. If G has n vertices, the game lasts at most n rounds. In Round i, where 1 < i < n,
Builder presents the vertex vi of G and describes all edges joining vi with vertices in {vj 1 < j < i}. This information is complete and correct. In particular, if the game lasts all n rounds, then Builder must have correctly :
specified the entire graph. After receiving the information for the new vertex vi, Colorer must then assign to vi a color from the set {1, 2, ... , t } so that this color is distinct from those previously assigned to neighbors of vi. These assignments are permanent. The (t, G) game ends at Round i and Builder is the winner if Colorer has no legitimate choice of a color for the new vertex vi. If on the other hand, Colorer is able to respond with a legitimate color for each of the n vertices of G, then Colorer is the winner. The on-line chromatic number of a graph G
is then the least t for which Colorer has a winning strategy for the (t, G) game-regardless of the strategy employed by Builder. In [71], Kierstead and Trotter prove the following foundational result.
Theorem 14.1 The on-line chromatic number of an interval graph of maximum clique size k is at most 3k - 2.
Proof Here's the winning strategy for Colorer. Given a new vertex x by Builder, Colorer assigns x to a set Si where i is the least positive integer for which there is no complete subgraph of size i + 1 containing x and i other vertices previously assigned to Sl U S2 U .. U Si. Note that S1 is just an independent set, so it can be colored with a single color. For each i > 2, we show that First Fit will color Si with the 3 colors from the set {3i - 4, 3i - 3, 3i - 2}. We accomplish this by showing that for i > 2, Si is the disjoint sum of paths. Fix i > 2. When a vertex u is presented by Builder and assigned by Colorer to Si, as opposed to Si_1, there is a clique Ku consisting of u and i - 1 vertices U Si_1. Then the intersection of the intervals from Ku is a from S1 U S2 U nonempty interval Iu, which is contained in the interval corresponding to u.
Now let u and v be adjacent vertices from Si. Then it is easy to see that Iu does not intersect the interval corresponding to v and I does not intersect the interval corresponding to u. From this, it follows easily that Si is triangle free and that each vertex from Si has at most 2 neighbors in Si. The algorithm presented in the preceding theorem has one feature in common with the First Fit algorithm discussed in Theorem 4.2: it is not necessary to know the maximum clique size in advance. If First Fit is used to color an interval graph, and the vertices are not processed in the order of left end points,
262
William T. Trotter
then it is not clear how many colors will be used. In [65], Kierstead showed that First Fit will use at most 40k colors on an interval graph with maximum clique size k, regardless of the order in which the vertices are processed. Subsequently, Kierstead and Qin [70] improved this upper bound to 26k. From below, Chrobak and Slusarek [23] showed that no on-line algorithm can color all interval graphs with maximum clique size k with fewer than 4.4k colors. Kierstead's analysis of the performance of First Fit in coloring interval graphs provided a solution to an important long standing problem in computer science called the Dynamic Storage Allocation problem. The standard two dimensional bin packing problem is to pack a family of rectangles in 1R2, with sides parallel to the coordinate axes, into a region of minimum area. The Dynamic Storage Allocation problem is to pack the rectangles into a region of minimum height-when the projections of the rectangles onto the horizontal coordinate axis form a fixed interval graph. Of course, by "pack," we mean that the rectangles are to be placed so that their interiors are disjoint. So if the maximum sum of the heights of rectangles whose projections have a common point is t, then t is a lower bound on the height required for a packing, and it was conjectured that a height of 0(t) would suffice. One proposed approach to finding a reasonably good packing was to assume all rectangles had height a power of 2. This assumption would at most double the optimal height required for a packing. These rectangles would then be partitioned into subrectangles of height one. Finally, First Fit would be used to color the rectangles (intervals) with all intervals formed from the same rectangle colored consecutively. The number of colors used by First Fit would then be an upper bound on the minimum height required for a packing. Accordingly, Kierstead and Qin's bound implies that the rectangles can in fact be packed into a region of height 52t. We refer the reader to [66] for a full discussion. As an added bonus, this paper provides an alternative approach to the dynamic storage allocation problem which stands as the best solution to date. This approach uses the same partition of rectangles, but colors them with a modified version of the on-line algorithm used in Theorem 14.1 rather than with First Fit. The end result is to show that the rectangles can be packed into a height of 6t - 4. Ironically, the research which led to the proof of Theorem 14.1 was motivated, not by the Dynamic Storage Allocation problem, but by the on-line version of Dilworth's theorem. In [64], Kierstead proved that there is an on-line algorithm which will partition a poset built one point at a time into (5w -1)/4 chains, where w denotes the width of the poset. When the poset is known to be an interval order, then the preceding theorem asserts that 3w - 2 chains suffice.
Kierstead's on-line chain partitioning algorithm requires knowledge of the order. Just knowing whether points are comparable is not enough. However, for interval graphs, our algorithm only makes use of the comparability graph. For many years, it remained an open problem to determine whether a com-
Interval Orders and Interval Graphs
263
parability graph of independence number k can be partitioned on-line into a bounded number of complete subgraphs. An affirmative answer was provided by Kierstead, Penrice and Trotter in [69]. In [68], Kierstead, McNulty and Trotter investigate on-line dimension. Here the game is between a Realizer and a Builder, with Realizer building a family R of linear extensions of a poset P which Builder is constructing one point at a time. They show that the on-line dimension of a class of width 4 posets is infinite. However, the posets in this class all contain the 3-dimensional
crown S. They then proceed to show that the on-line dimension of posets of bounded width is well defined, provided that the posets are crown-free, i.e., do not contain any 3-dimensional crown S.
Theorem 14.2 The on-line dimension of a crown-free poset of width k is at most t!, where t = (5k+1 - 1)/4. On the surface, this result has nothing to do with interval orders, but the proof makes use of an auxiliary order at a critical point in the argument. This structure turns out to be an interval order, and the order-theoretic properties this structure gains from being an interval order are key elements of the proof. Other sources of information about on-line coloring include [72] and the more recent survey by Kierstead [67]. In particular, this last paper contains a concise treatment of the recent breakthrough where Kierstead succeeded in showing that for all k > 3, there exists an e > 0 so that the on-line chromatic number of any k-colorable graph on n vertices is at most n1 . Kierstead's argument shows that e = 0(1/k!). Probably, this can be improved to 0(1/k). Another good source of problems (some of which are on-line) concerning interval graphs and other classes of perfect graphs in Gyarfas' survey paper [54]. In [27], Faigle, Kern, Kierstead and Trotter consider the following game theoretic problem for graphs. Two players, Alice and Bob, color a graph G using elements of the set [t] as colors. They alternate turns with Alice having the first move. Alice wins if the graph is eventually colored and Bob (an uncooperative partner) wins if at some step before the graph is colored, there
is no legitimate move. The game chromatic number of G is the least t for which Alice has a winning strategy. For example, it is shown in [27] that the game chromatic number of a tree is at most 4; furthermore, this result is best possible. In [74], Kierstead and Trotter show that a planar graph has game chromatic number at most 33; they also show that there exists a planar graph with game chromatic number at least 8. For interval graphs, the following result, given in [27], provides the best known bound on the game chromatic number of an interval graph.
Theorem 14.3 The game chromatic number of an interval graph G = (V, E) with maximum clique size k is at most 3k - 2.
Proof Let I be a distinguishing representation of an interval graph G with maximum clique size k. When it is her turn to color, Alice prefers to color a
William T. Trotter
264
vertex x adjacent to the vertex just colored by Bob. Among such, she prefers those whose intervals intersect the interval corresponding to the vertex just colored by Bob. Finally among such vertices, Alice prefers the one whose interval has right end point as large as possible. She then colors this vertex by
First Fit. We claim that Alice and Bob can never reach an impasse if the number of colors is 3k - 2. It suffices to show that the strategy given for Alice can be used by either player. Let x be the vertex to be colored. It suffices to show that x has at most 3k - 3 colored neighbors. Split the colored neighbors into three sets N1 U N2 U N3, where
1. N1 is the set of colored neigbors of x whose intervals contain the right end point of I (x); 2. N2 is the set of colored neighbors of x whose intervals are properly contained in I (x); and
3. N3 is the set of colored neighbors of x whose intervals contain the left end point of I (x) but not the right.
Clearly, INil < k - 1 and IN31 < k - 1, so our claim follows if we can show that IN21 < k - 1. Now our strategy for Alice insures that she will not have colored any of the vertices in N2, since she will always prefer to color x. So all vertices in N2 are colored by Bob, and at every turn-except possibly the last
one-Alice has selected a vertex other than x to color. Such a vertex must have an interval containing the interval corresponding to the vertex in N2 just
colored by Bob, and its right end point is greater than the right end point of x. Therefore Alice's response was to color a vertex from N1. It follows that IN2I < k. Now suppose that IN21 = k. Among the vertices in N2, let y be the unique vertex whose right end point is as large as possible. Then y, x and the k - 1 vertices in N2 form a clique of size k + 1.
The reader should note that it is just a coincidence that the expression 3k - 2 appears in both the preceding two theorems. In the first case, we know that it is best possible, but in the second, we believe it is not. We leave it as an exercise to show that, for each k > 2, there exists an interval graph G whose maximum clique size is k and whose game chromatic number is at least 2k.
15
Fractional dimension and ramsey theory for probability spaces
It is often useful to consider a fractional version of an integer valued combinatorial parameter as, in many cases, the resulting LP relaxation sheds light on the original problem. In [20], Brightwell and Scheinerman proposed to investigate fractional dimension for posets.
Interval Orders and Interval Graphs
265
Let P = (X, P) be a poset, and let F = {M1,. .. , Mt} be a multiset of linear extensions of P. Brightwell and Scheinerman [20] call F a k -fold realizer of P if, for each incomparable pair (x, y), there are at least k linear extensions in F which reverse the pair (x, y), i.e.,{i : 1 < i < t, x > y in Mz}I > k. The fractional dimension of P, denoted by fdim(P), is then defined as the least real number q > 1 for which there exists a k-fold realizer .F = {M1, . . . , Mt} of P so that k/t > 1/q (it is easily verified that the least upper bound of such real numbers q is indeed attained and is therefore a rational number). Using this terminology, the dimension of P is just the least t for which there exists a 1-fold realizer of P. It follows immediately that fdim(P) < dim(P), for every
poset P. The dimension or fractional dimension of a class of posets is defined to be
the least upper bound of dim(P) (respectively fdim(P)) over all posets P in the class. We have seen that dim(Z) = oo for the class Z of interval orders, but Brightwell and Scheinerman showed that fdim(Z) < 4. To see this, observe that if P = (X, P) is an interval order and A C X, there is a linear extension L of P with x > y in L for any incomparable pair (x, y) with x E A and y ¢ A. Building a realizer from one such L for each subset A of X of size LIXI/2] gives fdim(P) < 4. Brightwell and Scheinerman conjectured in [20] that fdim(Z) = 4, even though no example of an interval order of fractional dimension even as high as 3 was then known. In the remainder of this section, we sketch the approach taken by Trotter and Winkler in [126] to settle this conjecture in the affirmative. First, the following preliminary result is required. Intuitively, this theorem asserts that in a sufficiently long sequence of events, one cannot do substantially better than toss a fair coin in trying to balance between events being true and events being false. See [126] for the proof.
Theorem 15.1 For every c > 0, there exists an integer mo so that if m > mo and {Uz : 1 < i < m} is any sequence of events in a probability space, then there exist integers i and j with 1 < i < j < m so that Prob[UUUj] < 1+E.
With this result in hand, we can now sketch the proof of the solution. The argument makes extensive use of ramsey theory to make certain statements about sets hold in a uniform manner. To be precise, these statements involve small errors, and the argument takes some care to show that the errors can be kept under control. In this sketch, we ignore these errors.
Theorem 15.2 For every e > 0, there exists an integer no so that if n > no, the fractional dimension of the canonical interval order In is at least 4 - E.
Proof Let e > 0, and suppose that fdim(In) < 4 - E, regardless of the size of n. We argue to a contradiction, provided n is sufficiently large. Let S = {s1, s2, ... , 52m} be a 2m-element subset of [n], with sl < s2
[si, sm,+i]. Using ramsey theory, it is relatively easy to see that for fixed m, if n is sufficiently large, we may assume that the probability of U(S) is constant, for all 2m-element subsets of [n]. But Trotter and Winkler show more. They show that one may also assume that the event U(S) depends only on sl and sm+1. We denote this event by U(x, y), where x = s1 and y = sm+1.
Dually, let D(S) denote the event that for some j with 1 < j < in, [si, Sm+j] > [sm, 82m]. This time, the event D(S) depends only on sm and S2m. So we can just write D(x, y), where x = sm and y = S2m It follows that one can find a large homogeneous subset H so that U(x, y) fl D(x, y) = 0, for every x, y E H with x < y in R. Furthermore, if x < y < z < w in H, then U(x, z) fl U(y, w) = 0. If the homogenous set H has more than 2mo elements, the result follows from Theorem 15.1.
The dimension problem for interval orders is closely related to the graph coloring problem for shift graphs, a subject of independent interest. Similarly, the research on the fractional dimension of interval orders has led to many new and interesting concepts. We give hints to one of these in the sketch of the proof of Theorem 15.2, namely the development of a general Ramsey theory for probability spaces. However, there are several concrete combinatorial problems which are also quite attractive. Fix integers n and k with 1 < k < n. Suppose we have a probability space containing an event Es for every k-element subset S of [n]. We abuse notation
and just refer to this event as S. Now consider the minimum probability Prob(AB) taken over all (k, n)-shift pairs. In turn, take the maximum value of this probability over all probability spaces and let n go to infinity. The resulting value is called f (k). For example, from Theorem 15.1, it follows that
f (1) = 4. In [126], Trotter and Winkler prove that f (2) = 1, f (3) > s and f (4) > 5. In general, they prove that f (k) is strictly increasing and converges to 2. The relaxation of dimension to fractional dimension is an appealing concept.
In [33], Felsner and Trotter show that the fractional dimension of a poset in which each point is comparable with at most k others is at most k + 1. They also prove several other inequalities linking fractional dimension with width and cardinality. Nevertheless, there are many challenging open questions in this area.
16
Higher-dimensional analogues for graphs
In recent years, there has been a steady stream of results providing higher dimension analogues of interval graphs. Perhaps the first of these is due to Roberts [96] who defined the boxicity of a graph G = (V, E) as the least t for which there exists a function B assigning to each vertex x E V a sequence (II(1), I(2),. .. , I,,(t)) of closed intervals of 1R so that {x, y} E E if and only if Ii(i) fl ly(i) # 0, for all i E [t]. Equivalently, the boxicity of a graph is just
Interval Orders and Interval Graphs
267
the least t so that the graph is the intersection of boxes in R. So interval graphs are graphs with boxicity one. Roberts showed that the boxicity of a graph on n vertices is at most [n/2j, when n > 2. For example the graph Hn, obtained by taking the complement of a matching on n edges has 2n vertices and boxicity n. In [127], Wittenshausen showed that for all n > 1, Hn is the only graph with 2n vertices and boxicity n. However, when a graph has 2n + 1 vertices and boxicity n, the situation is modestly more complicated. For example, the cycle C5 on 5 vertices has boxicity 2. Also, the graph W3 with vertex set 11, 2, ... , 7} and edges joining i to i + 2, i + 3, i + 4 and i + 5 (cyclically) has boxicity 3. In [113], Trotter showed that a graph G on 2n + 1 vertices has boxicity n if and only if one of the following conditions holds:
1. G contains H. 2. Gn contains the join of C5 and Hn_2. 3. Gn contains the join of W3 and Hn_3. In [109], Thomassen showed that the boxicity of a planar graph is at most 3; in fact, the boxes corresponding to adjacent vertices may be required to intersect on a face. Many of the basic concepts for interval graphs have natural interpretation for digraphs. In [3], Beineke and Zamfirescu introduced the notion (with dif-
ferent terminology) of an interval digraph. By this we mean that for each vertex x in a digraph D, there are two intervals of the real line Rx and S.', so that D contains a directed arc from x to y if and only if RxnSy # 0. Structural questions for interval digraphs are studied in [84], [85], [105] and [106]. Define the interval number of a graph G = (V, E) as the least t for which G is the intersection graph of a family of sets, with each set being the union of t pairwise disjoint closed intervals of R. In [53], Griggs and West show that if the maximum degree of G is d, then the interval number of G is at most
[(d+1)/ni. This inequality is tight if G is triangle-free. Griggs and West also showed that there exists an absolute constant c > 0 so that the interval number of a graph with q edges is at most c,,Fq. In [103], Scheinerman and West showed that the interval number of a planar graph is at most 3, and Scheinerman [100] showed that there exists an absolute constant c' > 0 so that the interval number of a graph of genus ry is at most c' f. In [122], Trotter and Harary show that the interval number of a complete
bipartite graph K(m, n) is [(mn + 1)/(m + n)]. The fact that the interval number of K(m, n) is at least this large follows from the following elementary observation. The interval number of a triangle-free graph G with n vertices and q edges is at least [(q + 1)/nl. This inequality follows from the fact
that if the interval number of the graph is t, then the nt intervals used in a representation form the intersection graph of a forest on nt vertices and at least q edges. This requires q < nt - 1.
268
William T. Trotter
Somewhat surprisingly, the determination of the interval numbers for complete multipartite graphs proved to be more challenging. The interval number of the complete multipartite graph K(nl, n2i ... , ns), with nl > n2 > > nt > 2, is at least as large as the interval number of K(nl, n2). Call this quantity to. In [59], Hopkins, Trotter and West show the interval number of K(nl, n2,. . ., ns) is at most to + 1. Furthermore, they show that it is equal to to, except possibly for the two cases (ni, n2) = (7,5) and nl = n2 - n2 - 1. In both these exceptional cases, the interval number of K(nl, n2, ... , nt) may equal to + 1, provided there are enough other parts of appropriate size. Motivated by the formula for complete bipartite graphs, Trotter and Harary [122] conjectured that the maximum interval number of a graph on n vertices is 1(n + 1)/41. This conjecture was proved by Griggs in [52]. In [22], Chang and West introduced the concept of interval number for digraphs. For a digraph D, the interval number of D is just the least positive integer i for which there exists a function F assigning to each vertex x two subsets R,,, S,, of the real numbers so that 1. For each node x in D, Rx and S., are each the union of at most t pairwise disjoint intervals of R, and
2. D contains an arc from x to y if and only if R,,, fl Sy
0.
Chang and West showed that the maximum interval number of a digraph on it nodes is O(n/log n). They also defined the concept of boxicity for digraphs and showed that the maximum boxicity of a digraph on n nodes is In/2]. Aigner and Andreae [1] introduced an interesting variation of interval number. For an graph G = (V, E), they defined the total interval number of G as the least positive integer t for which there exists a function F assigning to each vertex x of G a set F(x) which is the union of t, pairwise disjoint closed intervals of IR so that: 1. For every x, y E V, {x, y} E E if and only if F(x) fl F(y)
0, and
2. >xEX tX = t.
Aigner and Andreae [1] produced upper bounds on total interval number for several classes of graphs. For example, they showed that the maximum total interval number of a tree on n nodes is L(5n - 3)/4j . In [80], Kratzke and West showed that the maximum total interval number of an outerplanar graph on it nodes is L3n/2 - 1J while the maximum total interval number of a general graph on it nodes is 1(n2 + 1)/41. These results settled conjectures made by Aigner and Andreae in [1]. Other results on total interval number are given by Kostochka and West in [79]; in particular, they bound the total interval number in terms of the maximum degree, and characterize graphs for which the bound is sharp. The components of these graphs are balanced complete bipartite graphs.
Interval Orders and Interval Graphs
269
In [81], Kratzke and West provide a linear time algorithm for computing the total interval number of a tree, and they show that it is NP-complete to test whether the total interval number of a graph is exactly one more than the number of edges, even for the class of triangle-free, 3-regular planar graphs.
Given a poset P = (X, P) and points x, y E X, with x < y in P, the interval [x, y] is just the set {u E X : x < u < y in P}. The poset boxicity of a graph G is the least t for which there exists a t-dimensional poset P for which G is the intersection graph of intervals in P. In [125], Trotter and West show that there exists an absolute constant c > 0 so that the poset boxicity of a graph on n vertices is at most c log log n. They also show that there exist graphs with arbitrarily large poset boxicity. In [56], Gyarfas and West consider the multitrack interval number of a graph as the least t for which the graph is the union of t interval orders. We will discuss analogous concepts for posets in Sections 17 and 19.
17
Higher dimensional analogues for orders The investigation of higher dimensional analogues of interval orders has also
produced a steady stream of results. First, let P be any hereditary class of orders which contains the linear orders. Then we can define the P-dimension of a poset P = (X, P) as the least t for which P is the intersection of t orders from P. The hereditary property serves to ensure that the P-dimension of P is at most the P-dimension of Q when P is contained in Q. Of course, the Pdimension of P is at most dim(P), and to emphasize the distinction between the original definition of dimension and variants discussed in the remainder of this paper, the dimension is also called the ordinary dimension. In [14], Bogart and Trotter defined the interval dimension of a poset P = (X, P) as the least t for which P is the intersection of t interval orders on X. So a poset has interval dimension 1 if and only if it is an interval order. Posets with interval dimension at most 2 have also been studied extensively. In [114], Trotter gave a forbidden subposet characterization of height two posets having interval dimension at most 2. This characterization results in a complete listing of all minimal posets of height 2 having interval dimension 3. Polynomial time
recognition algorithms for posets having interval dimension at most 2 have
been provided by several authors, but the best to date is due to Ma and Spinrad [86].
One of the most appealing aspects of interval dimension is the positive solution of the removable pair conjecture. For ordinary dimension, Trotter conjectured (see [118], for example) that if P is a poset having three or more points, then there is always a pair of points whose removal decreases the dimension by at most 1. In fact, he conjectured that the removal of a critical pair always decreases the dimension by at most 1. Although the removable pair conjecture remains open, this second conjecture was disproved by Reuter [95], and an infinite family of counterexamples was then constructed by Kierstead
William T. Trotter
270
and Trotter [73]. However, for interval dimension, we have the following elementary result.
Theorem 17.1 Let P = (X, P) be a poset and let (x, y) E crit(P). If Q = (Y, Q) is the subposet determined by Y = X - {x, y}, then the interval dimension of P is at most one more than the interval dimension of Q. Proof Let Q1i Q2i ... , Qt be interval orders on Y whose intersection is Q. For
each i E [t], let Pi be an interval order on X so that Pi(Y) = Qi. Then let L be any linear extension of Y with D(x) < Y - D(x) and Y - U(y) < U(y) in L . Define a partial order Pt+1 on X by setting Pt+1 = P U L. It is easy to see that Pt+1 is an interval order and that P = P1 fl P2 fl . f1 Pt+1
Another appealing aspect of the concept of interval dimension is that there is a relatively simple characterization of posets having maximal dimension for a given number of points (see Bogart and Trotter [13]), while the corresponding problem for ordinary dimension is considerably more difficult. Several other inequalities relating interval dimension to other combinatorial parameters are simpler than the corresponding results for ordinary dimension, e.g., compare the forbidden subposet characterization of the inequality dim(P, X) < max{2, IX - Al}, when A is an antichain, for ordinary dimension [111] with the result for interval dimension in [13]. Other aspects of the interplay between dimension and interval dimension are discussed in [30]. In [57], Habib, Kelly and Mohring show that the property of a poset having interval dimension at most 2 is a comparability invariant, i.e., it depends only on the underlying comparability graph and not on the specific order. Bogart and Trotter also defined the semi-order dimension of a poset and noted that if the semi-order dimension of P is t, then the ordinary dimension of P is at most 3t. This result is tight when t = 1, but it is not known whether it is best possible when t > 2. In [31], Felsner and Mohring show that the property of a poset having semi-order dimension at most 2 is a comparability invariant. In a somewhat different direction, more closely connected to the concepts discussed in the preceding section, Madej and West [87] define the interval inclusion number of a poset P = (X, P) as the least integer t for which there exist a function F assigning to each x E X a set F(x) C R so that:
1. For each x E X, F(x) is the union of at most t pairwise disjoint closed intervals of IR, and
2. For each x, y E X, x < y in P if and only if F(x) 9 F(y). In [88], Madej and West show that "almost all" posets on n points have interval
number o(n), but it is still not known whether there exists a positive real number c so that for all n, there exists a poset on n points with interval
Interval Orders and Interval Graphs
271
inclusion number exceeding cn. It is easy to see that the interval inclusion number of an n-dimensional poset is at most [n/21. Furthermore, as Madej and West note in [88], the set of all subsets of an n-element set, ordered by inclusion, shows that this last inequality is tight. However, as they point out, the n-dimensional standard example S, has interval inclusion number 2, for
alln>2. 18
Intervals, angles and spheres
Over the past 10 years, there has been a flurry of work on geometric problems which arise when posets are represented by a family of sets (usually some geometrically defined objects) ordered by inclusion. These structures are called
inclusion orders, and they are the natural order theoretic analogue of intersection graphs. For example, as is well known, a poset has dimension at most 2
if and only if it isomorphic to the inclusion order determined by a family of intervals of the real line. Space limitations do not allow us to discuss the full range of research on inclusion orders, but we will attempt to highlight those which related directly to interval orders. Fishburn and Trotter [40] define a poset P = (X, P) to be an angle order when P is the inclusion order of a family of subsets of the euclidean plane, with each set being an angular region determined by two rays emanating from a common point. They show that every interval order is an angle order and that every poset with dimension at most 4 is an angle order. They also showed that there exists a 7-dimensional poset which is not an angle order. Subsequently, several authors showed that there exists a 5-dimensional poset which is not an angle order, but the most elegant proof of this fact results from the theory of "degrees of freedom" developed by Alon and Scheinerman in [2]. A d-sphere with center x and radius r is the set of all points in Rd whose distance to x is at most r. A 1-sphere is just a closed interval of IR. Call a poset
P a sphere order if there is some d so that P is isormorphic to the inclusion order determined by a family of d-spheres in W. When P is a sphere order, we may define the sphere dimension of a poset P = (X, P) as the least d for which P is the inclusion order of a family of d-spheres. So a poset has sphere dimension 1 if and only if it has ordinary dimension at most 2. The problem of determining whether every finite poset is a sphere order is posed by Brightwell and Winkler in [21], and it is widely believed that the answer is negative. When d = 2, there are some interesting results and one especially vexing problem. For historical reasons, posets with sphere dimension at most 2 are called circle orders, although it might have been more accurate to call them disk orders. The recent article [101] contains a number of interesting perspectives
on the problem of representing order by circles in the plane. The range and extent of the connections with other combinatorial problems is most surprising. In [37], Fishburn shows that every interval order is a circle order. Trivially,
William T. Trotter
272
every poset with ordinary dimension at most 2 is a circle order-in fact, we can require that the circles all have centers on a fixed line in the plane. By the Alon/Scheinerman theory, there exist 4-dimensional posets which are not circle orders. However, it is not known whether every finite 3-dimensional poset is a circle order. Scheinerman and Weirman [102] showed that the countably infinite 3-dimensional poset N3 is not a circle order. Subsequently, a somewhat shorter proof of this result was given by Hurlbert [60]. The sharpest result to
date is due to Fon-der-Flaass [43] who showed that 2 x 3 x N is not a sphere order, but that 2 x 2 x N is a circle order. On the other hand, it is an easy exercise to show that if P is a finite poset with ordinary dimension at most 3 and n > 3, then P is the inclusion order of a family of regular n-gons in the euclidean plane, and it is easy to suspect that when n is quite large relative to IXI, these polygons are extremely close to being circles. However, I would conjecture that there is a finite 3-dimensional poset which is not a sphere order. In this discussion, the metric used to determine distance plays a critical role. Of course, if x = (x1, x2, ... , xd) and y = (yl, y2,. .., yd), then the ordinary distance from x to y is Ea 1(x, - y;,)2. But if we change this to max{Jxi - yzJ 1 < i < d}, then a d-sphere is just a cube. Furthermore, it is an easy exercise to show that every poset with dimension at most d + 1 is the inclusion order of a family of cubes in V. Again, by the Alon/Scheinerman theory, this is best possible, meaning that there are (d + 2)-dimensional posets which cannot be represented by cubes in W ordered by inclusion. :
19
Tolerances, thresholds and gaps
In the preceding two sections, we discussed higher dimensional analogues for interval graphs and interval orders. In this section, we discuss generalizations which arise when just one interval is assigned but more complex rules are used to determine edges and comparabilities. Here is the basic motivation. If we have an indexed family.F = {I (x) : x E X} of closed intervals with distinct end points, then an interval graph results when we define an edge set E by {x, y} E X if and only if 1I(x) n I(y) I > 0. From an applications standpoint, the problem with this definition is that we take two vertices to be adjacent when their intervals intersect regardless of how small this intersection might be. Similarly, an interval order assigns x to be less than y only when F(x) lies entirely to the left on F(y). But there are many scheduling problems where we want to consider one job as preceding another even when there is some overlap in time. We begin with generalizations of interval graphs. Golumbic and Monma [49] proposed the following definition. Given an indexed family .F = {I (x) : x E X}
of closed intervals of R and a subset T = {tx x E X1 of the non-negative real numbers 180, define the tolerance graph G = G(F,T) = (X, E) by setting
Interval Orders and Interval Graphs
273
E = {{x, y} : x, y E X, x y and 1I(x) f1I(y)I > min{t,,ty}}. It is easy to see that an interval graph is a tolerance graph. Just take a distinguishing representation and give each vertex a tolerance smaller than the distance between any two end points used in the representation. Also, the complement of an interval graph is a tolerance graph. In this case, for each x E X, set tx _ II(x)I. A tolerance graph is bounded if 0 < tx < II(x)I, for all x E X. Golumbic and Monma [49] showed that a bounded tolerance graph is the complement of a comparability graph and is therefore perfect. This argument does not work for tolerance graphs which are not bounded, but in [50], Golumbic, Monma and Trotter showed that all tolerance graphs are perfect. The proof in the general case follows by showing that the complement of a tolerance graph is perfectly orderable. Note that Chvatal's concept of a graph being perfectly orderable [24] is a weakening of the key property used to show that interval graphs are perfect. Here is an interesting way in which tolerance graphs differ from interval graphs. Recall that an interval graph is proper if and only if it has a represent-
ation using only intervals of length 1. This is not true for tolerance graphs. In [6], Bogart, Fishburn, Isaak and Langley show that the class of proper tolerance graphs is strictly larger than the class of tolerance graphs in which all intervals have unit length. In the last several years, a number of new concepts for generalizing tolerance
graphs have been introduced. Perhaps the most general is due to Jacobson, McMorris and Mulder [61] who proposed to study graphs defined by a family of intervals {I,, : x E X}, a subset T = {tom : x E X} of tolerances drawn from the set R0 of non-negative reals, and a function 0: ][8o x 1R0 -> Ro by setting the edge set E to consist of all 2-element sets {x, y} for which 1I(x) f1 I(y) > ty).
The original definition of a tolerance graph is just the function 0(t", ty) _ min{t,,, ty}.
Now here are some of the new ideas for posets. McMorris and Jacobsen (see [8]) propose to study a generalization of interval orders in which extra conditions are imposed on the gaps between intervals corresponding to comparable pairs of points. The definition requires an indexed family {I(x) = [a.,, b,,] : x E X} of closed intervals of R, a subset T = {t,: : x E X} of the non-negative reals ][80 i and a function 0:180 x 1180 -3 R0 . We then define a relation P on X by setting (x, y) E P if and only if (1) x = y or ty). We call these posets 0-gap orders to reflect that the (2) by - a, > relation P is defined in terms of the gap between the two intervals.
In certain cases, P will be a partial order on X. For example, this is always the case if 0 satisfies the triangle inequality: 0(tx, ty) + ¢(ty, tz) >
0(t,,, t,), for all t,, ty, tz E R. In particular, P is always a partial order if 0(tx, ty) = max{tx, ty}. On the other hand, we may fail to get a partial order if 0(t, ty) = min{tx, ty}. The special case where 0(t,,, ty) = max{t,,, ty} is called a max-gap order.
In another direction, Bogart and Trenk [12] call a poset P = (X, P) a
William T. Trotter
274
bi-tolerance order when there exists a triple (I, F, G) where: 1. I assigns to each x E X a closed interval I (x) = [ax, bx] of R;
2. F={fx:xEX}CR,G={gx:xEX}CR; 3. ax < fx, gx < bx in R, for every x E X;
4. x < y in P if and only if bx < f, and g A2 > ... > AN > -1 where N is the number of states, and all these eigenvalues are real. The rate of convergence to the distribution 7r is determined by the quantity .Amax = max{A2,
IaNI}.
More precisely we have the following proposition. Let P2j(t) denote
Pr{Xt+h =
I Xh = 2}
Approximate Counting
295
and let
i. (t) =
lPij(t) - ir(j)I
2
denote what is called the variation distance at time t. It is clearly a measure of the separation from the stationary distribution at time t. Now define, for c > 0, the mixing time function Ti by Ti(e) = min{t : Ai(s) < e
Vs > t}.
The following result shows the relationship between mixing times and the maximum eigenvalues. It is an extension by Sinclair [69] of a key result from [70].
Proposition 4.1 For e > 0, Ti(e) satisfies 1
1
(1) Ti (e) < 1 - Amax (In Sr(i) + In e );
(ii) max r (e) > i
2(1 mAmax) In ( 2e) .
In order to achieve rapid convergence we need ri (e) be small, for all i. It is useful to note the following trick which concentrates the interest on A2. Replace P by P' = (I + P) where I is the identity matrix. This only affects rates of convergences by a polynomial factor. All eigenvalues of P' are nonnegative, and the quantity Amax of P' is !Amax, so that henceforth we need only consider the second eigenvalue A2. Ideally we want 1 - A2 to be large so A2 must be small. The key idea of Sinclair and Jerrum [70] was to relate this to the very aptly named conductance 4D of the chain. This is defined by
= max sc_cz
E P(i, j)7r(i)
7r(i) iES
iES,jEQ\S
But the bracketed term is just the conditional probability of leaving a set S in the equilibrium state. In other words I is a measure of the ability of the chain to escape from any subset of the state space Q. Theorem 4.2 The second eigenvalue A2 of a reversible ergodic Markov chain satisfies
1-21